Allanatrix commited on
Commit
97ced61
·
verified ·
1 Parent(s): 920e102

Upload NexaMass V3 safetensors weights and release metadata

Browse files
README.md CHANGED
@@ -30,7 +30,7 @@ The model was developed as part of the Nexa MS/MS pipeline. The foundation check
30
 
31
  ## Model Summary
32
 
33
- `NexaMass-V3-Struct` is an encoder-only transformer with approximately `12.38M` trainable parameters. The checkpoint is approximately `153 MB`, making it substantially smaller than many public MS/MS representation or chemistry foundation checkpoints. The model uses `256` peak tokens, an input feature width of `8`, hidden dimension `384`, `6` transformer layers, `8` attention heads, feed-forward dimension `1536`, dropout `0.1`, SSL projection dimension `192`, RDKit Morgan fingerprint dimension `2048`, target projection dimension `256`, and retrieval MLP hidden dimension `512`.
34
 
35
  The model contains a spectral transformer backbone, an SSL projection head, a Morgan fingerprint structure head, a spectrum-side retrieval query projection, a fingerprint-side target projection, and experimental retrieval/reranking heads. The strongest current inference surface is the Morgan fingerprint prediction and candidate-bank decode path. The trained retrieval projection is included for research, but it is not promoted as a reliable final decision layer.
36
 
@@ -40,6 +40,9 @@ Use this model to embed MS/MS spectra, build nearest-neighbor or clustering anal
40
 
41
  Do not use this model as an unrestricted de novo structure generator. Do not treat the top-ranked candidate as a production-grade molecular identification without external validation. Do not treat the raw confidence or score gap as calibrated probability. Candidate outputs should be validated with chemistry-aware checks such as RDKit sanitization, formula and exact-mass agreement, precursor/adduct compatibility, fingerprint similarity, and spectrum-to-candidate consistency.
42
 
 
 
 
43
  ## Inference Contract
44
 
45
  The recommended inference path is spectrum-to-embedding plus candidate-bank decoding. A spectrum is converted into a normalized embedding and a `2048`-dimensional Morgan fingerprint probability vector. That vector is compared against candidate molecular fingerprints from a known bank. The output should be a ranked shortlist with validation metadata, not a single unqualified structure claim.
@@ -60,9 +63,9 @@ The structure-aligned checkpoint, `NexaMass-V3-Struct`, resumed from the foundat
60
 
61
  ## Evaluation
62
 
63
- The structure-aligned checkpoint reached validation structure BCE `0.0653`, validation fingerprint cosine `0.4255`, and embedding standard deviation `0.0722` in the V26 alignment run. A candidate-bank structure gallery over `20` displayed validation examples recovered `11` exact fingerprint-decoded molecular identities through the fingerprint decode path. The same gallery recovered `0` exact identities through the trained retrieval projection, which is why the card treats ranking as an open decision-layer problem rather than a solved property.
64
 
65
- On a MassSpecGym retrieval adapter using the frozen V3 projected-dot scorer, the model reached validation Hit@1 `0.1162`, Hit@5 `0.1915`, and Hit@20 `0.3328`. On the MassSpecGym test dataloader under Hit@k-only evaluation, it reached Hit@1 `0.0627`, Hit@5 `0.1753`, and Hit@20 `0.3505`. This places the model above lower MassSpecGym retrieval baselines such as Random, DeepSets, Fingerprint FFN, and DeepSets with Fourier features, but below specialized retrieval systems such as MIST. The result should be interpreted as evidence of useful top-k narrowing, not leaderboard-grade retrieval.
66
 
67
  Internal larger-bank pressure tests also support a bounded interpretation. On a roughly `29k` unique-identity candidate bank, the frozen projected-dot scorer beat random and compressed candidate rank substantially, but exact local discrimination remained weak. The best current read is that the model contains real structure and retrieval signal, while top-1 ranking and calibrated confidence require additional downstream machinery.
68
 
@@ -78,7 +81,7 @@ MS/MS structure inference can affect downstream scientific interpretation. Users
78
 
79
  ## Citation
80
 
81
- If you use this model, cite the NexaMass project release and the accompanying technical report or paper draft when available. Relevant background work includes DreaMS for self-supervised MS/MS representation learning, MassSpecGym for benchmark framing, CSI:FingerID for fingerprint-mediated candidate search, and related spectra-structure retrieval and de novo generation systems such as MIST, MSNovelist, CMSSP, CSU-MS2, MSBERT, Spec2Mol, and MS2Mol.
82
 
83
  ## Recommended Name
84
 
 
30
 
31
  ## Model Summary
32
 
33
+ `NexaMass-V3-Struct` is an encoder-only transformer with `14,106,690` trainable parameters. The full training checkpoint is approximately `153 MB` in object storage, while the public model-state weights are approximately `54 MB` per checkpoint and are provided in both PyTorch and Safetensors formats, making the released model substantially smaller than many public MS/MS representation or chemistry foundation checkpoints. The model uses `256` peak tokens, an input feature width of `8`, hidden dimension `384`, `6` transformer layers, `8` attention heads, feed-forward dimension `1536`, dropout `0.1`, SSL projection dimension `192`, RDKit Morgan fingerprint dimension `2048`, target projection dimension `256`, and retrieval MLP hidden dimension `512`.
34
 
35
  The model contains a spectral transformer backbone, an SSL projection head, a Morgan fingerprint structure head, a spectrum-side retrieval query projection, a fingerprint-side target projection, and experimental retrieval/reranking heads. The strongest current inference surface is the Morgan fingerprint prediction and candidate-bank decode path. The trained retrieval projection is included for research, but it is not promoted as a reliable final decision layer.
36
 
 
40
 
41
  Do not use this model as an unrestricted de novo structure generator. Do not treat the top-ranked candidate as a production-grade molecular identification without external validation. Do not treat the raw confidence or score gap as calibrated probability. Candidate outputs should be validated with chemistry-aware checks such as RDKit sanitization, formula and exact-mass agreement, precursor/adduct compatibility, fingerprint similarity, and spectrum-to-candidate consistency.
42
 
43
+
44
+ The preferred public inference files are `weights/NexaMass-V3-Struct-model_state.safetensors` and `weights/Final_V3-model_state.safetensors`. PyTorch model-state `.pt` fallbacks are included for compatibility. Optimizer-bearing full training checkpoints are retained in Wasabi/object storage and are not part of the public Hugging Face payload.
45
+
46
  ## Inference Contract
47
 
48
  The recommended inference path is spectrum-to-embedding plus candidate-bank decoding. A spectrum is converted into a normalized embedding and a `2048`-dimensional Morgan fingerprint probability vector. That vector is compared against candidate molecular fingerprints from a known bank. The output should be a ranked shortlist with validation metadata, not a single unqualified structure claim.
 
63
 
64
  ## Evaluation
65
 
66
+ The structure-aligned checkpoint reached validation structure BCE `0.0653`, validation fingerprint cosine `0.4255`, and embedding standard deviation `0.0722` in the V26 alignment run. A candidate-bank structure gallery over `20` displayed validation examples matched the ground-truth candidate identity in `11` cases through the fingerprint decode path. The same gallery matched `0` exact identities through the trained retrieval projection, which is why the card treats ranking as an open decision-layer problem rather than a solved property.
67
 
68
+ On a MassSpecGym retrieval adapter using the frozen V3 projected-dot scorer, the model reached validation Hit@1 `0.1162`, Hit@5 `0.1915`, and Hit@20 `0.3328`. On the MassSpecGym test dataloader under Hit@k-only evaluation, it reached Hit@1 `0.0627`, Hit@5 `0.1753`, and Hit@20 `0.3505`. This Hit@20-only adapter comparison places the model above lower MassSpecGym retrieval baselines such as Random, DeepSets, Fingerprint FFN, and DeepSets with Fourier features, but below specialized retrieval systems such as MIST. The result should be interpreted as evidence of useful top-k narrowing, not leaderboard-grade retrieval.
69
 
70
  Internal larger-bank pressure tests also support a bounded interpretation. On a roughly `29k` unique-identity candidate bank, the frozen projected-dot scorer beat random and compressed candidate rank substantially, but exact local discrimination remained weak. The best current read is that the model contains real structure and retrieval signal, while top-1 ranking and calibrated confidence require additional downstream machinery.
71
 
 
81
 
82
  ## Citation
83
 
84
+ If you use this model, cite the NexaMass project release and the accompanying technical report when available. Relevant background work includes DreaMS for self-supervised MS/MS representation learning, MassSpecGym for benchmark framing, CSI:FingerID for fingerprint-mediated candidate search, and related spectra-structure retrieval and de novo generation systems such as MIST, MSNovelist, CMSSP, CSU-MS2, MSBERT, Spec2Mol, and MS2Mol.
85
 
86
  ## Recommended Name
87
 
config.json ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architecture_config": {
3
+ "collision_max": 200.0,
4
+ "dropout": 0.1,
5
+ "fingerprint_dim": 2048,
6
+ "heads": 8,
7
+ "layers": 6,
8
+ "max_peaks": 256,
9
+ "metadata_scale": 0.02,
10
+ "model_dim": 384,
11
+ "mz_max": 2000.0,
12
+ "projection_dim": 192,
13
+ "retrieval_mlp_hidden_dim": 512,
14
+ "target_projection_dim": 256
15
+ },
16
+ "architectures": [
17
+ "NexaMassSpectralEncoder"
18
+ ],
19
+ "foundation_checkpoint": "weights/Final_V3-model_state.safetensors",
20
+ "foundation_checkpoint_format": "safetensors",
21
+ "foundation_pytorch_model_state_checkpoint": "weights/Final_V3-model_state.pt",
22
+ "full_training_checkpoints": {
23
+ "location": "Wasabi object storage",
24
+ "note": "Optimizer-bearing full training checkpoints are intentionally not part of the public Hugging Face payload; the public release contains model-state weights only."
25
+ },
26
+ "input_contract": {
27
+ "feature_width": 8,
28
+ "mask": "boolean valid-peak mask, shape [batch, max_peaks]",
29
+ "max_peaks": 256,
30
+ "metadata_embeddings": [
31
+ "adduct_id",
32
+ "instrument_id"
33
+ ],
34
+ "per_peak_features": [
35
+ "mz",
36
+ "intensity",
37
+ "mz_to_precursor",
38
+ "peak_rank",
39
+ "precursor_mz",
40
+ "charge",
41
+ "collision_energy",
42
+ "peak_count"
43
+ ]
44
+ },
45
+ "library_name": "pytorch",
46
+ "model_type": "nexamass",
47
+ "not_supported": [
48
+ "unrestricted de novo molecule generation",
49
+ "calibrated top-1 molecular identification without external validation"
50
+ ],
51
+ "outputs": {
52
+ "morgan_fingerprint_logits": 2048,
53
+ "retrieval_target_dim": 256,
54
+ "ssl_projection_dim": 192,
55
+ "structure_query_dim": 256
56
+ },
57
+ "parameter_count": 14106690,
58
+ "primary_checkpoint": "weights/NexaMass-V3-Struct-model_state.safetensors",
59
+ "primary_checkpoint_format": "safetensors",
60
+ "pytorch_model_state_checkpoint": "weights/NexaMass-V3-Struct-model_state.pt",
61
+ "recommended_inference": "Use the structure-aligned model-state checkpoint for spectrum embeddings and Morgan fingerprint probability vectors, then compare against a candidate molecular fingerprint bank.",
62
+ "task": "MS/MS spectral representation learning and RDKit Morgan fingerprint alignment"
63
+ }
configs/foundation_config.json ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint": {
3
+ "bytes": 151419519,
4
+ "config": {
5
+ "collision_max": 200.0,
6
+ "dropout": 0.1,
7
+ "fingerprint_dim": 2048,
8
+ "heads": 8,
9
+ "layers": 6,
10
+ "max_peaks": 256,
11
+ "metadata_scale": 0.02,
12
+ "model_dim": 384,
13
+ "mz_max": 2000.0,
14
+ "projection_dim": 192,
15
+ "retrieval_mlp_hidden_dim": 512,
16
+ "target_projection_dim": 256
17
+ },
18
+ "created_at": "2026-04-13T19:01:18Z",
19
+ "file": "weights/Final_V3.pt",
20
+ "metrics_keys": [
21
+ "backbone_lr",
22
+ "contrastive_loss",
23
+ "covariance_loss",
24
+ "covariance_to_contrastive_ratio",
25
+ "covariance_weight",
26
+ "dataloader_time_sec",
27
+ "dataloader_wait_pct_300s",
28
+ "dataloader_wait_pct_60s",
29
+ "dataloader_wait_pct_900s",
30
+ "dataloader_wait_pct_run",
31
+ "distributed_rank",
32
+ "distributed_world_size",
33
+ "effective_batch_size",
34
+ "effective_hard_negative_scale",
35
+ "effective_retrieval_weight",
36
+ "effective_structure_bce_weight",
37
+ "embedding_std",
38
+ "global_batch_size",
39
+ "gpu_active_fraction",
40
+ "gpu_count_observed",
41
+ "gpu_max_memory_gb",
42
+ "gpu_mem_used_mean_gb",
43
+ "gpu_mem_util_mean_pct",
44
+ "gpu_memory_utilization",
45
+ "gpu_snapshot_available",
46
+ "gpu_util_max_pct",
47
+ "gpu_util_mean_pct",
48
+ "gpu_util_min_pct",
49
+ "grad_norm",
50
+ "local_batch_size",
51
+ "lr",
52
+ "mfu_estimate",
53
+ "msms_objective",
54
+ "optimizer_step",
55
+ "regularizer_ramp",
56
+ "resumed_from_checkpoint",
57
+ "retrieval_head_lr",
58
+ "samples_per_sec",
59
+ "samples_per_sec_300s",
60
+ "samples_per_sec_60s",
61
+ "samples_per_sec_900s",
62
+ "samples_per_sec_run_mean",
63
+ "step_time_cv_300s",
64
+ "step_time_cv_60s",
65
+ "step_time_cv_900s",
66
+ "step_time_sec",
67
+ "structure_bce_locked",
68
+ "structure_bce_loss",
69
+ "structure_fingerprint_cosine",
70
+ "structure_head_lr",
71
+ "structure_retrieval_loss",
72
+ "structure_retrieval_skipped",
73
+ "structure_target_backend",
74
+ "structure_to_ssl_ratio",
75
+ "train_loss",
76
+ "train_loss_ema",
77
+ "train_retrieval_bank_size",
78
+ "variance_loss",
79
+ "variance_weight",
80
+ "weighted_covariance_loss",
81
+ "weighted_regularizer_to_contrastive_ratio",
82
+ "weighted_ssl_loss",
83
+ "weighted_structure_bce_loss",
84
+ "weighted_structure_retrieval_loss",
85
+ "weighted_variance_loss"
86
+ ],
87
+ "model_state_bytes": 56473018,
88
+ "model_state_file": "weights/Final_V3-model_state.pt",
89
+ "param_count": 14106690,
90
+ "state_tensors": 103,
91
+ "step": 81001
92
+ },
93
+ "dataset": "GeMS-derived deterministic Parquet shard campaign",
94
+ "name": "NexaMass-V3 foundation checkpoint",
95
+ "objective_summary": "self-supervised spectral contrastive representation learning with structure-facing foundation instrumentation",
96
+ "phase": "self_supervised_ssl_structure_foundation",
97
+ "processed_samples_approx": 201000000
98
+ }
configs/runtime_requirements.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "core_packages": [
3
+ "torch",
4
+ "numpy",
5
+ "pyarrow",
6
+ "rdkit"
7
+ ],
8
+ "framework": "PyTorch",
9
+ "load_example": [
10
+ "payload = torch.load(\"weights/NexaMass-V3-Struct-model_state.pt\", map_location=\"cpu\")",
11
+ "cfg = ModelConfig(**payload[\"args-derived config or configs/config.json\"])",
12
+ "model = SpectralEncoder(cfg)",
13
+ "model.load_state_dict(payload[\"model_state\"], strict=True)"
14
+ ],
15
+ "optional_packages": [
16
+ "qdrant-client",
17
+ "huggingface-hub"
18
+ ],
19
+ "python": ">=3.11",
20
+ "reference_runtime": "Nexa_compute scripts.msms.inspect_msms_checkpoint_structure.SpectralEncoder"
21
+ }
configs/structure_alignment_config.json ADDED
@@ -0,0 +1,162 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "checkpoint": {
3
+ "bytes": 160482975,
4
+ "config": {
5
+ "collision_max": 200.0,
6
+ "dropout": 0.1,
7
+ "fingerprint_dim": 2048,
8
+ "heads": 8,
9
+ "layers": 6,
10
+ "max_peaks": 256,
11
+ "metadata_scale": 0.02,
12
+ "model_dim": 384,
13
+ "mz_max": 2000.0,
14
+ "projection_dim": 192,
15
+ "retrieval_mlp_hidden_dim": 512,
16
+ "target_projection_dim": 256
17
+ },
18
+ "created_at": "2026-04-13T21:34:38Z",
19
+ "file": "weights/NexaMass-V3-Struct.pt",
20
+ "metrics_keys": [
21
+ "backbone_lr",
22
+ "contrastive_loss",
23
+ "covariance_loss",
24
+ "covariance_to_contrastive_ratio",
25
+ "covariance_weight",
26
+ "dataloader_time_sec",
27
+ "dataloader_wait_pct_300s",
28
+ "dataloader_wait_pct_60s",
29
+ "dataloader_wait_pct_900s",
30
+ "dataloader_wait_pct_run",
31
+ "distributed_rank",
32
+ "distributed_world_size",
33
+ "effective_batch_size",
34
+ "effective_hard_negative_scale",
35
+ "effective_retrieval_weight",
36
+ "effective_structure_bce_weight",
37
+ "embedding_std",
38
+ "global_batch_size",
39
+ "gpu_active_fraction",
40
+ "gpu_count_observed",
41
+ "gpu_max_memory_gb",
42
+ "gpu_mem_used_mean_gb",
43
+ "gpu_mem_util_mean_pct",
44
+ "gpu_memory_utilization",
45
+ "gpu_snapshot_available",
46
+ "gpu_util_max_pct",
47
+ "gpu_util_mean_pct",
48
+ "gpu_util_min_pct",
49
+ "grad_norm",
50
+ "local_batch_size",
51
+ "lr",
52
+ "mfu_estimate",
53
+ "msms_objective",
54
+ "optimizer_step",
55
+ "regularizer_ramp",
56
+ "resumed_from_checkpoint",
57
+ "retrieval_head_lr",
58
+ "samples_per_sec",
59
+ "samples_per_sec_300s",
60
+ "samples_per_sec_60s",
61
+ "samples_per_sec_900s",
62
+ "samples_per_sec_run_mean",
63
+ "step_time_cv_300s",
64
+ "step_time_cv_60s",
65
+ "step_time_cv_900s",
66
+ "step_time_sec",
67
+ "structure_bce_locked",
68
+ "structure_bce_loss",
69
+ "structure_candidate_count",
70
+ "structure_clear_fraction",
71
+ "structure_clear_hit_at_1",
72
+ "structure_close_fraction",
73
+ "structure_close_hit_at_1",
74
+ "structure_confidence_bucket_score_mean",
75
+ "structure_confidence_bucket_score_mode",
76
+ "structure_confidence_candidate_count",
77
+ "structure_confidence_clear_threshold",
78
+ "structure_confidence_close_threshold",
79
+ "structure_confidence_eval_key_mode",
80
+ "structure_confidence_separation",
81
+ "structure_correct_logit_mean",
82
+ "structure_fingerprint_cosine",
83
+ "structure_hard_negative_bonus_applied",
84
+ "structure_hard_negative_bonus_mean",
85
+ "structure_hard_negative_scale",
86
+ "structure_head_lr",
87
+ "structure_hit_at_1",
88
+ "structure_hit_at_20",
89
+ "structure_hit_at_5",
90
+ "structure_local_ce_loss",
91
+ "structure_local_ce_loss_weight",
92
+ "structure_local_rerank_loss",
93
+ "structure_local_rerank_loss_weight",
94
+ "structure_local_rerank_score_scale",
95
+ "structure_logit_margin_mean",
96
+ "structure_margin_loss",
97
+ "structure_margin_loss_weight",
98
+ "structure_missing_formula_fraction",
99
+ "structure_missing_inchikey_fraction",
100
+ "structure_pairwise_logistic_loss",
101
+ "structure_pairwise_logistic_loss_weight",
102
+ "structure_positive_count_mean",
103
+ "structure_post_label_candidate_count",
104
+ "structure_post_precursor_candidate_count",
105
+ "structure_prefilter_candidate_count",
106
+ "structure_retrieval_logit_std",
107
+ "structure_retrieval_loss",
108
+ "structure_retrieval_scorer",
109
+ "structure_retrieval_target_mode",
110
+ "structure_skipped_query_fraction",
111
+ "structure_soft_target_exact_mix",
112
+ "structure_soft_target_loss",
113
+ "structure_soft_target_loss_weight",
114
+ "structure_target_backend",
115
+ "structure_target_equivalent_count_mean",
116
+ "structure_target_identity_collision_count",
117
+ "structure_target_rank_mean",
118
+ "structure_to_ssl_ratio",
119
+ "structure_top1_top2_gap_mean",
120
+ "structure_top1_top2_gap_norm_mean",
121
+ "structure_top_negative_logit_mean",
122
+ "structure_uncertain_fraction",
123
+ "structure_uncertain_hit_at_1",
124
+ "structure_valid_query_fraction",
125
+ "train_loss",
126
+ "train_loss_ema",
127
+ "train_retrieval_bank_size",
128
+ "variance_loss",
129
+ "variance_weight",
130
+ "weighted_covariance_loss",
131
+ "weighted_regularizer_to_contrastive_ratio",
132
+ "weighted_ssl_loss",
133
+ "weighted_structure_bce_loss",
134
+ "weighted_structure_retrieval_loss",
135
+ "weighted_variance_loss"
136
+ ],
137
+ "model_state_bytes": 56474172,
138
+ "model_state_file": "weights/NexaMass-V3-Struct-model_state.pt",
139
+ "param_count": 14106690,
140
+ "state_tensors": 103,
141
+ "step": 91001
142
+ },
143
+ "dataset": "corrected labeled GeMS handoff surface with enforced molecular labels",
144
+ "evaluation_summary": {
145
+ "embedding_std": 0.0722,
146
+ "gallery_exact_fingerprint_decodes": 11,
147
+ "gallery_size": 20,
148
+ "massspecgym_test_hit20_adapter": 0.3505,
149
+ "validation_fingerprint_cosine": 0.4255,
150
+ "validation_structure_bce": 0.0653
151
+ },
152
+ "name": "NexaMass-V3-Struct alignment checkpoint",
153
+ "objective_weights": {
154
+ "retrieval_local_ce_loss_weight": 0.4,
155
+ "retrieval_soft_target_loss_weight": 0.3,
156
+ "retrieval_weight": 0.6,
157
+ "ssl_preservation_weight": 0.05,
158
+ "structure_bce_weight": 1.0
159
+ },
160
+ "parent_checkpoint": "weights/Final_V3-model_state.pt",
161
+ "phase": "RDKit Morgan fingerprint structure alignment"
162
+ }
metadata/Final_V3_manifest.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model_name": "Final_V3",
3
+ "created_at_utc": "2026-04-13T19:53:04Z",
4
+ "source_checkpoint": "/home/ubuntu/artifacts/nexa_mol/v1_full_pretrain/phase1_blackwell_v3/current_checkpoint.pt",
5
+ "checkpoint_alias": "/home/ubuntu/artifacts/nexa_mol/v1_full_pretrain/phase1_blackwell_v3/Final_V3.pt",
6
+ "bytes": 151419519,
7
+ "run_name": "gems_v1_full_pretrain_blackwell_phase1_v3",
8
+ "dataset_root": "/home/ubuntu/data/gems_v1_v1_full_pretrain_phase1_v3",
9
+ "phase": "ssl_structure_foundation_final_v3"
10
+ }
metadata/archive/20260413T201817Z/latest_checkpoint.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ /home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt
metadata/archive/20260413T202042Z/latest_checkpoint.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ /home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt
metadata/archive/20260413T213902Z/day1_validation_report.json ADDED
@@ -0,0 +1,2475 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "analysis_basis": [
3
+ "docs/projects/msms_pipeline/Technical_Plan.md",
4
+ "docs/projects/msms_pipeline/Engineering.md"
5
+ ],
6
+ "created_at": "2026-04-13T21:36:19Z",
7
+ "dataset_root": "/home/ubuntu/data/gems_v1_handoff_labeled_v26",
8
+ "debug_metric_pressures": {
9
+ "gpu_max_memory_gb": {
10
+ "desired_direction": "stable",
11
+ "early_mean": 13.252925169950275,
12
+ "improvement_pct": 100.0,
13
+ "improvement_pct_clipped": 100.0,
14
+ "last": 13.279157161712646,
15
+ "late_mean": 13.279157161712646,
16
+ "mid_mean": 13.279059082097401,
17
+ "pressure": "pressure_up",
18
+ "pressure_arrow": "\u2197",
19
+ "pressure_score": 1.0
20
+ },
21
+ "structure_confidence_separation": {
22
+ "desired_direction": "higher",
23
+ "early_mean": -0.007918493679557749,
24
+ "improvement_pct": -531.5617836245987,
25
+ "improvement_pct_clipped": -100.0,
26
+ "last": -0.0040241447277367115,
27
+ "late_mean": -0.05001017991881603,
28
+ "mid_mean": -0.032136323077151974,
29
+ "pressure": "pressure_down",
30
+ "pressure_arrow": "\u2198",
31
+ "pressure_score": -1.0
32
+ },
33
+ "structure_correct_logit_mean": {
34
+ "desired_direction": "higher",
35
+ "early_mean": 6.134031791907514,
36
+ "improvement_pct": 47.305071917778086,
37
+ "improvement_pct_clipped": 47.305071917778086,
38
+ "last": 9.1875,
39
+ "late_mean": 9.035739942528735,
40
+ "mid_mean": 9.474891618497109,
41
+ "pressure": "pressure_up",
42
+ "pressure_arrow": "\u2197",
43
+ "pressure_score": 1.0
44
+ },
45
+ "structure_fingerprint_cosine": {
46
+ "desired_direction": "higher",
47
+ "early_mean": 0.3965519580968543,
48
+ "improvement_pct": 17.58747890372286,
49
+ "improvement_pct_clipped": 17.58747890372286,
50
+ "last": 0.46021854877471924,
51
+ "late_mean": 0.46629545006943846,
52
+ "mid_mean": 0.4464927057310336,
53
+ "pressure": "pressure_up",
54
+ "pressure_arrow": "\u2197",
55
+ "pressure_score": 1.0
56
+ },
57
+ "structure_hit_at_1": {
58
+ "desired_direction": "higher",
59
+ "early_mean": 0.007880238439306358,
60
+ "improvement_pct": 549.6830023383724,
61
+ "improvement_pct_clipped": 100.0,
62
+ "last": 0.00390625,
63
+ "late_mean": 0.051196569683908046,
64
+ "mid_mean": 0.032706376445086706,
65
+ "pressure": "pressure_up",
66
+ "pressure_arrow": "\u2197",
67
+ "pressure_score": 1.0
68
+ },
69
+ "structure_hit_at_20": {
70
+ "desired_direction": "higher",
71
+ "early_mean": 0.10721865968208093,
72
+ "improvement_pct": 239.4614028144399,
73
+ "improvement_pct_clipped": 100.0,
74
+ "last": 0.1171875,
75
+ "late_mean": 0.3639659662356322,
76
+ "mid_mean": 0.28105807442196534,
77
+ "pressure": "pressure_up",
78
+ "pressure_arrow": "\u2197",
79
+ "pressure_score": 1.0
80
+ },
81
+ "structure_hit_at_5": {
82
+ "desired_direction": "higher",
83
+ "early_mean": 0.03360955563583815,
84
+ "improvement_pct": 402.47008675709174,
85
+ "improvement_pct_clipped": 100.0,
86
+ "last": 0.029296875,
87
+ "late_mean": 0.16887796336206898,
88
+ "mid_mean": 0.11673591040462428,
89
+ "pressure": "pressure_up",
90
+ "pressure_arrow": "\u2197",
91
+ "pressure_score": 1.0
92
+ },
93
+ "structure_logit_margin_mean": {
94
+ "desired_direction": "higher",
95
+ "early_mean": -1.1263999277456647,
96
+ "improvement_pct": 1.2274364862230593,
97
+ "improvement_pct_clipped": 1.2274364862230593,
98
+ "last": -1.0078125,
99
+ "late_mean": -1.1125740840517242,
100
+ "mid_mean": -1.0154725885115607,
101
+ "pressure": "pressure_none",
102
+ "pressure_arrow": "\u2192",
103
+ "pressure_score": 0.0
104
+ },
105
+ "structure_retrieval_logit_std": {
106
+ "desired_direction": "higher",
107
+ "early_mean": 1.0455416592559375,
108
+ "improvement_pct": 112.05988926255111,
109
+ "improvement_pct_clipped": 100.0,
110
+ "last": 0.9839304089546204,
111
+ "late_mean": 2.2171744848119803,
112
+ "mid_mean": 2.0036371215919537,
113
+ "pressure": "pressure_up",
114
+ "pressure_arrow": "\u2197",
115
+ "pressure_score": 1.0
116
+ },
117
+ "structure_retrieval_loss": {
118
+ "desired_direction": "lower",
119
+ "early_mean": 10.237479978903181,
120
+ "improvement_pct": 10.094066619063353,
121
+ "improvement_pct_clipped": 10.094066619063353,
122
+ "last": 10.630162239074707,
123
+ "late_mean": 9.204101929719421,
124
+ "mid_mean": 9.41523380775672,
125
+ "pressure": "pressure_up",
126
+ "pressure_arrow": "\u2197",
127
+ "pressure_score": 1.0
128
+ },
129
+ "structure_target_rank_mean": {
130
+ "desired_direction": "lower",
131
+ "early_mean": 351.84337743858384,
132
+ "improvement_pct": 58.80171764880722,
133
+ "improvement_pct_clipped": 58.80171764880722,
134
+ "last": 314.703125,
135
+ "late_mean": 144.9534280711207,
136
+ "mid_mean": 176.16389315390174,
137
+ "pressure": "pressure_up",
138
+ "pressure_arrow": "\u2197",
139
+ "pressure_score": 1.0
140
+ },
141
+ "structure_top1_top2_gap_mean": {
142
+ "desired_direction": "higher",
143
+ "early_mean": 0.09793567106213873,
144
+ "improvement_pct": 59.221149173532794,
145
+ "improvement_pct_clipped": 59.221149173532794,
146
+ "last": 0.0986328125,
147
+ "late_mean": 0.1559343009159483,
148
+ "mid_mean": 0.12284365968208093,
149
+ "pressure": "pressure_up",
150
+ "pressure_arrow": "\u2197",
151
+ "pressure_score": 1.0
152
+ },
153
+ "structure_top1_top2_gap_norm_mean": {
154
+ "desired_direction": "higher",
155
+ "early_mean": 0.12215662799301864,
156
+ "improvement_pct": -35.653274040957065,
157
+ "improvement_pct_clipped": -35.653274040957065,
158
+ "last": 0.10155132412910461,
159
+ "late_mean": 0.07860379065547524,
160
+ "mid_mean": 0.07032911504236604,
161
+ "pressure": "pressure_down",
162
+ "pressure_arrow": "\u2198",
163
+ "pressure_score": -1.0
164
+ },
165
+ "structure_top_negative_logit_mean": {
166
+ "desired_direction": "lower",
167
+ "early_mean": 7.262102601156069,
168
+ "improvement_pct": -39.724147080005764,
169
+ "improvement_pct_clipped": -39.724147080005764,
170
+ "last": 10.1875,
171
+ "late_mean": 10.14691091954023,
172
+ "mid_mean": 10.490606936416185,
173
+ "pressure": "pressure_down",
174
+ "pressure_arrow": "\u2198",
175
+ "pressure_score": -1.0
176
+ },
177
+ "structure_uncertain_fraction": {
178
+ "desired_direction": "lower",
179
+ "early_mean": 0.924663565751445,
180
+ "improvement_pct": -6.108024336610137,
181
+ "improvement_pct_clipped": -6.108024336610137,
182
+ "last": 0.970703125,
183
+ "late_mean": 0.9811422413793104,
184
+ "mid_mean": 0.984747561416185,
185
+ "pressure": "pressure_down",
186
+ "pressure_arrow": "\u2198",
187
+ "pressure_score": -1.0
188
+ },
189
+ "train_loss_ema": {
190
+ "desired_direction": "lower",
191
+ "early_mean": 6.265581717362126,
192
+ "improvement_pct": 10.455707205061195,
193
+ "improvement_pct_clipped": 10.455707205061195,
194
+ "last": 6.519052214060582,
195
+ "late_mean": 5.610470838300897,
196
+ "mid_mean": 5.8371378655061745,
197
+ "pressure": "pressure_up",
198
+ "pressure_arrow": "\u2197",
199
+ "pressure_score": 1.0
200
+ },
201
+ "val_ssl_loss": {
202
+ "desired_direction": "lower",
203
+ "early_mean": 1.0849772497183745,
204
+ "improvement_pct": 2.3192742421198114,
205
+ "improvement_pct_clipped": 2.3192742421198114,
206
+ "last": 1.0397093277424574,
207
+ "late_mean": 1.0598136518327963,
208
+ "mid_mean": 1.0747967428483425,
209
+ "pressure": "pressure_none",
210
+ "pressure_arrow": "\u2192",
211
+ "pressure_score": 0.0
212
+ },
213
+ "val_structure_correct_logit_mean": {
214
+ "desired_direction": "higher",
215
+ "early_mean": 8.0311279296875,
216
+ "improvement_pct": 16.44145855816145,
217
+ "improvement_pct_clipped": 16.44145855816145,
218
+ "last": 9.49462890625,
219
+ "late_mean": 9.3515625,
220
+ "mid_mean": 8.441162109375,
221
+ "pressure": "pressure_up",
222
+ "pressure_arrow": "\u2197",
223
+ "pressure_score": 1.0
224
+ },
225
+ "val_structure_retrieval_logit_std": {
226
+ "desired_direction": "higher",
227
+ "early_mean": 1.2326001720502973,
228
+ "improvement_pct": 102.32641409365677,
229
+ "improvement_pct_clipped": 100.0,
230
+ "last": 2.6622845763340592,
231
+ "late_mean": 2.49387572822161,
232
+ "mid_mean": 2.022283688886091,
233
+ "pressure": "pressure_up",
234
+ "pressure_arrow": "\u2197",
235
+ "pressure_score": 1.0
236
+ },
237
+ "val_structure_retrieval_loss": {
238
+ "desired_direction": "lower",
239
+ "early_mean": 11.562767315655947,
240
+ "improvement_pct": 4.864921250164783,
241
+ "improvement_pct_clipped": 4.864921250164783,
242
+ "last": 10.882395941764116,
243
+ "late_mean": 11.000247791409492,
244
+ "mid_mean": 12.150514367967844,
245
+ "pressure": "pressure_none",
246
+ "pressure_arrow": "\u2192",
247
+ "pressure_score": 0.0
248
+ },
249
+ "val_structure_top1_top2_gap_norm_mean": {
250
+ "desired_direction": "higher",
251
+ "early_mean": 0.06950851557485294,
252
+ "improvement_pct": -16.826381575968767,
253
+ "improvement_pct_clipped": -16.826381575968767,
254
+ "last": 0.05411380449368153,
255
+ "late_mean": 0.0578127475164365,
256
+ "mid_mean": 0.07577372609375743,
257
+ "pressure": "pressure_down",
258
+ "pressure_arrow": "\u2198",
259
+ "pressure_score": -1.0
260
+ },
261
+ "val_structure_top_negative_logit_mean": {
262
+ "desired_direction": "lower",
263
+ "early_mean": 9.2255859375,
264
+ "improvement_pct": -16.07653223245475,
265
+ "improvement_pct_clipped": -16.07653223245475,
266
+ "last": 10.714599609375,
267
+ "late_mean": 10.708740234375,
268
+ "mid_mean": 10.51171875,
269
+ "pressure": "pressure_down",
270
+ "pressure_arrow": "\u2198",
271
+ "pressure_score": -1.0
272
+ },
273
+ "weighted_regularizer_to_contrastive_ratio": {
274
+ "desired_direction": "lower",
275
+ "early_mean": 0.0667222202939108,
276
+ "improvement_pct": 54.60656419810926,
277
+ "improvement_pct_clipped": 54.60656419810926,
278
+ "last": 0.028996589606316913,
279
+ "late_mean": 0.03028750823471251,
280
+ "mid_mean": 0.0329437448262181,
281
+ "pressure": "pressure_up",
282
+ "pressure_arrow": "\u2197",
283
+ "pressure_score": 1.0
284
+ }
285
+ },
286
+ "debug_metric_pressures_display": {
287
+ "gpu_max_memory_gb": {
288
+ "desired_direction": "stable",
289
+ "early_mean": 13.252925169950275,
290
+ "goal_display": "goal \u2248",
291
+ "improvement_pct": 100.0,
292
+ "improvement_pct_display": "+100.00%",
293
+ "last": 13.279157161712646,
294
+ "late_mean": 13.279157161712646,
295
+ "mid_mean": 13.279059082097401,
296
+ "pressure": "pressure_up",
297
+ "pressure_arrow": "\u2197",
298
+ "pressure_display": "goal \u2248 | \u2197 +100.00% improving",
299
+ "pressure_score": 1.0,
300
+ "trend_label": "improving"
301
+ },
302
+ "structure_confidence_separation": {
303
+ "desired_direction": "higher",
304
+ "early_mean": -0.007918493679557749,
305
+ "goal_display": "goal \u2191",
306
+ "improvement_pct": -531.5617836245987,
307
+ "improvement_pct_clipped": -100.0,
308
+ "improvement_pct_clipped_display": "-100.00%",
309
+ "improvement_pct_display": "-531.56%",
310
+ "last": -0.0040241447277367115,
311
+ "late_mean": -0.05001017991881603,
312
+ "mid_mean": -0.032136323077151974,
313
+ "pressure": "pressure_down",
314
+ "pressure_arrow": "\u2198",
315
+ "pressure_display": "goal \u2191 | \u2198 -531.56% worsening",
316
+ "pressure_score": -1.0,
317
+ "trend_label": "worsening"
318
+ },
319
+ "structure_correct_logit_mean": {
320
+ "desired_direction": "higher",
321
+ "early_mean": 6.134031791907514,
322
+ "goal_display": "goal \u2191",
323
+ "improvement_pct": 47.305071917778086,
324
+ "improvement_pct_display": "+47.31%",
325
+ "last": 9.1875,
326
+ "late_mean": 9.035739942528735,
327
+ "mid_mean": 9.474891618497109,
328
+ "pressure": "pressure_up",
329
+ "pressure_arrow": "\u2197",
330
+ "pressure_display": "goal \u2191 | \u2197 +47.31% improving",
331
+ "pressure_score": 1.0,
332
+ "trend_label": "improving"
333
+ },
334
+ "structure_fingerprint_cosine": {
335
+ "desired_direction": "higher",
336
+ "early_mean": 0.3965519580968543,
337
+ "goal_display": "goal \u2191",
338
+ "improvement_pct": 17.58747890372286,
339
+ "improvement_pct_display": "+17.59%",
340
+ "last": 0.46021854877471924,
341
+ "late_mean": 0.46629545006943846,
342
+ "mid_mean": 0.4464927057310336,
343
+ "pressure": "pressure_up",
344
+ "pressure_arrow": "\u2197",
345
+ "pressure_display": "goal \u2191 | \u2197 +17.59% improving",
346
+ "pressure_score": 1.0,
347
+ "trend_label": "improving"
348
+ },
349
+ "structure_hit_at_1": {
350
+ "desired_direction": "higher",
351
+ "early_mean": 0.007880238439306358,
352
+ "goal_display": "goal \u2191",
353
+ "improvement_pct": 549.6830023383724,
354
+ "improvement_pct_clipped": 100.0,
355
+ "improvement_pct_clipped_display": "+100.00%",
356
+ "improvement_pct_display": "+549.68%",
357
+ "last": 0.00390625,
358
+ "late_mean": 0.051196569683908046,
359
+ "mid_mean": 0.032706376445086706,
360
+ "pressure": "pressure_up",
361
+ "pressure_arrow": "\u2197",
362
+ "pressure_display": "goal \u2191 | \u2197 +549.68% improving",
363
+ "pressure_score": 1.0,
364
+ "trend_label": "improving"
365
+ },
366
+ "structure_hit_at_20": {
367
+ "desired_direction": "higher",
368
+ "early_mean": 0.10721865968208093,
369
+ "goal_display": "goal \u2191",
370
+ "improvement_pct": 239.4614028144399,
371
+ "improvement_pct_clipped": 100.0,
372
+ "improvement_pct_clipped_display": "+100.00%",
373
+ "improvement_pct_display": "+239.46%",
374
+ "last": 0.1171875,
375
+ "late_mean": 0.3639659662356322,
376
+ "mid_mean": 0.28105807442196534,
377
+ "pressure": "pressure_up",
378
+ "pressure_arrow": "\u2197",
379
+ "pressure_display": "goal \u2191 | \u2197 +239.46% improving",
380
+ "pressure_score": 1.0,
381
+ "trend_label": "improving"
382
+ },
383
+ "structure_hit_at_5": {
384
+ "desired_direction": "higher",
385
+ "early_mean": 0.03360955563583815,
386
+ "goal_display": "goal \u2191",
387
+ "improvement_pct": 402.47008675709174,
388
+ "improvement_pct_clipped": 100.0,
389
+ "improvement_pct_clipped_display": "+100.00%",
390
+ "improvement_pct_display": "+402.47%",
391
+ "last": 0.029296875,
392
+ "late_mean": 0.16887796336206898,
393
+ "mid_mean": 0.11673591040462428,
394
+ "pressure": "pressure_up",
395
+ "pressure_arrow": "\u2197",
396
+ "pressure_display": "goal \u2191 | \u2197 +402.47% improving",
397
+ "pressure_score": 1.0,
398
+ "trend_label": "improving"
399
+ },
400
+ "structure_logit_margin_mean": {
401
+ "desired_direction": "higher",
402
+ "early_mean": -1.1263999277456647,
403
+ "goal_display": "goal \u2191",
404
+ "improvement_pct": 1.2274364862230593,
405
+ "improvement_pct_display": "+1.23%",
406
+ "last": -1.0078125,
407
+ "late_mean": -1.1125740840517242,
408
+ "mid_mean": -1.0154725885115607,
409
+ "pressure": "pressure_none",
410
+ "pressure_arrow": "\u2192",
411
+ "pressure_display": "goal \u2191 | \u2192 +1.23% flat",
412
+ "trend_label": "flat"
413
+ },
414
+ "structure_retrieval_logit_std": {
415
+ "desired_direction": "higher",
416
+ "early_mean": 1.0455416592559375,
417
+ "goal_display": "goal \u2191",
418
+ "improvement_pct": 112.05988926255111,
419
+ "improvement_pct_clipped": 100.0,
420
+ "improvement_pct_clipped_display": "+100.00%",
421
+ "improvement_pct_display": "+112.06%",
422
+ "last": 0.9839304089546204,
423
+ "late_mean": 2.2171744848119803,
424
+ "mid_mean": 2.0036371215919537,
425
+ "pressure": "pressure_up",
426
+ "pressure_arrow": "\u2197",
427
+ "pressure_display": "goal \u2191 | \u2197 +112.06% improving",
428
+ "pressure_score": 1.0,
429
+ "trend_label": "improving"
430
+ },
431
+ "structure_retrieval_loss": {
432
+ "desired_direction": "lower",
433
+ "early_mean": 10.237479978903181,
434
+ "goal_display": "goal \u2193",
435
+ "improvement_pct": 10.094066619063353,
436
+ "improvement_pct_display": "+10.09%",
437
+ "last": 10.630162239074707,
438
+ "late_mean": 9.204101929719421,
439
+ "mid_mean": 9.41523380775672,
440
+ "pressure": "pressure_up",
441
+ "pressure_arrow": "\u2197",
442
+ "pressure_display": "goal \u2193 | \u2197 +10.09% improving",
443
+ "pressure_score": 1.0,
444
+ "trend_label": "improving"
445
+ },
446
+ "structure_target_rank_mean": {
447
+ "desired_direction": "lower",
448
+ "early_mean": 351.84337743858384,
449
+ "goal_display": "goal \u2193",
450
+ "improvement_pct": 58.80171764880722,
451
+ "improvement_pct_display": "+58.80%",
452
+ "last": 314.703125,
453
+ "late_mean": 144.9534280711207,
454
+ "mid_mean": 176.16389315390174,
455
+ "pressure": "pressure_up",
456
+ "pressure_arrow": "\u2197",
457
+ "pressure_display": "goal \u2193 | \u2197 +58.80% improving",
458
+ "pressure_score": 1.0,
459
+ "trend_label": "improving"
460
+ },
461
+ "structure_top1_top2_gap_mean": {
462
+ "desired_direction": "higher",
463
+ "early_mean": 0.09793567106213873,
464
+ "goal_display": "goal \u2191",
465
+ "improvement_pct": 59.221149173532794,
466
+ "improvement_pct_display": "+59.22%",
467
+ "last": 0.0986328125,
468
+ "late_mean": 0.1559343009159483,
469
+ "mid_mean": 0.12284365968208093,
470
+ "pressure": "pressure_up",
471
+ "pressure_arrow": "\u2197",
472
+ "pressure_display": "goal \u2191 | \u2197 +59.22% improving",
473
+ "pressure_score": 1.0,
474
+ "trend_label": "improving"
475
+ },
476
+ "structure_top1_top2_gap_norm_mean": {
477
+ "desired_direction": "higher",
478
+ "early_mean": 0.12215662799301864,
479
+ "goal_display": "goal \u2191",
480
+ "improvement_pct": -35.653274040957065,
481
+ "improvement_pct_display": "-35.65%",
482
+ "last": 0.10155132412910461,
483
+ "late_mean": 0.07860379065547524,
484
+ "mid_mean": 0.07032911504236604,
485
+ "pressure": "pressure_down",
486
+ "pressure_arrow": "\u2198",
487
+ "pressure_display": "goal \u2191 | \u2198 -35.65% worsening",
488
+ "pressure_score": -1.0,
489
+ "trend_label": "worsening"
490
+ },
491
+ "structure_top_negative_logit_mean": {
492
+ "desired_direction": "lower",
493
+ "early_mean": 7.262102601156069,
494
+ "goal_display": "goal \u2193",
495
+ "improvement_pct": -39.724147080005764,
496
+ "improvement_pct_display": "-39.72%",
497
+ "last": 10.1875,
498
+ "late_mean": 10.14691091954023,
499
+ "mid_mean": 10.490606936416185,
500
+ "pressure": "pressure_down",
501
+ "pressure_arrow": "\u2198",
502
+ "pressure_display": "goal \u2193 | \u2198 -39.72% worsening",
503
+ "pressure_score": -1.0,
504
+ "trend_label": "worsening"
505
+ },
506
+ "structure_uncertain_fraction": {
507
+ "desired_direction": "lower",
508
+ "early_mean": 0.924663565751445,
509
+ "goal_display": "goal \u2193",
510
+ "improvement_pct": -6.108024336610137,
511
+ "improvement_pct_display": "-6.11%",
512
+ "last": 0.970703125,
513
+ "late_mean": 0.9811422413793104,
514
+ "mid_mean": 0.984747561416185,
515
+ "pressure": "pressure_down",
516
+ "pressure_arrow": "\u2198",
517
+ "pressure_display": "goal \u2193 | \u2198 -6.11% worsening",
518
+ "pressure_score": -1.0,
519
+ "trend_label": "worsening"
520
+ },
521
+ "train_loss_ema": {
522
+ "desired_direction": "lower",
523
+ "early_mean": 6.265581717362126,
524
+ "goal_display": "goal \u2193",
525
+ "improvement_pct": 10.455707205061195,
526
+ "improvement_pct_display": "+10.46%",
527
+ "last": 6.519052214060582,
528
+ "late_mean": 5.610470838300897,
529
+ "mid_mean": 5.8371378655061745,
530
+ "pressure": "pressure_up",
531
+ "pressure_arrow": "\u2197",
532
+ "pressure_display": "goal \u2193 | \u2197 +10.46% improving",
533
+ "pressure_score": 1.0,
534
+ "trend_label": "improving"
535
+ },
536
+ "val_ssl_loss": {
537
+ "desired_direction": "lower",
538
+ "early_mean": 1.0849772497183745,
539
+ "goal_display": "goal \u2193",
540
+ "improvement_pct": 2.3192742421198114,
541
+ "improvement_pct_display": "+2.32%",
542
+ "last": 1.0397093277424574,
543
+ "late_mean": 1.0598136518327963,
544
+ "mid_mean": 1.0747967428483425,
545
+ "pressure": "pressure_none",
546
+ "pressure_arrow": "\u2192",
547
+ "pressure_display": "goal \u2193 | \u2192 +2.32% flat",
548
+ "trend_label": "flat"
549
+ },
550
+ "val_structure_correct_logit_mean": {
551
+ "desired_direction": "higher",
552
+ "early_mean": 8.0311279296875,
553
+ "goal_display": "goal \u2191",
554
+ "improvement_pct": 16.44145855816145,
555
+ "improvement_pct_display": "+16.44%",
556
+ "last": 9.49462890625,
557
+ "late_mean": 9.3515625,
558
+ "mid_mean": 8.441162109375,
559
+ "pressure": "pressure_up",
560
+ "pressure_arrow": "\u2197",
561
+ "pressure_display": "goal \u2191 | \u2197 +16.44% improving",
562
+ "pressure_score": 1.0,
563
+ "trend_label": "improving"
564
+ },
565
+ "val_structure_retrieval_logit_std": {
566
+ "desired_direction": "higher",
567
+ "early_mean": 1.2326001720502973,
568
+ "goal_display": "goal \u2191",
569
+ "improvement_pct": 102.32641409365677,
570
+ "improvement_pct_clipped": 100.0,
571
+ "improvement_pct_clipped_display": "+100.00%",
572
+ "improvement_pct_display": "+102.33%",
573
+ "last": 2.6622845763340592,
574
+ "late_mean": 2.49387572822161,
575
+ "mid_mean": 2.022283688886091,
576
+ "pressure": "pressure_up",
577
+ "pressure_arrow": "\u2197",
578
+ "pressure_display": "goal \u2191 | \u2197 +102.33% improving",
579
+ "pressure_score": 1.0,
580
+ "trend_label": "improving"
581
+ },
582
+ "val_structure_retrieval_loss": {
583
+ "desired_direction": "lower",
584
+ "early_mean": 11.562767315655947,
585
+ "goal_display": "goal \u2193",
586
+ "improvement_pct": 4.864921250164783,
587
+ "improvement_pct_display": "+4.86%",
588
+ "last": 10.882395941764116,
589
+ "late_mean": 11.000247791409492,
590
+ "mid_mean": 12.150514367967844,
591
+ "pressure": "pressure_none",
592
+ "pressure_arrow": "\u2192",
593
+ "pressure_display": "goal \u2193 | \u2192 +4.86% flat",
594
+ "trend_label": "flat"
595
+ },
596
+ "val_structure_top1_top2_gap_norm_mean": {
597
+ "desired_direction": "higher",
598
+ "early_mean": 0.06950851557485294,
599
+ "goal_display": "goal \u2191",
600
+ "improvement_pct": -16.826381575968767,
601
+ "improvement_pct_display": "-16.83%",
602
+ "last": 0.05411380449368153,
603
+ "late_mean": 0.0578127475164365,
604
+ "mid_mean": 0.07577372609375743,
605
+ "pressure": "pressure_down",
606
+ "pressure_arrow": "\u2198",
607
+ "pressure_display": "goal \u2191 | \u2198 -16.83% worsening",
608
+ "pressure_score": -1.0,
609
+ "trend_label": "worsening"
610
+ },
611
+ "val_structure_top_negative_logit_mean": {
612
+ "desired_direction": "lower",
613
+ "early_mean": 9.2255859375,
614
+ "goal_display": "goal \u2193",
615
+ "improvement_pct": -16.07653223245475,
616
+ "improvement_pct_display": "-16.08%",
617
+ "last": 10.714599609375,
618
+ "late_mean": 10.708740234375,
619
+ "mid_mean": 10.51171875,
620
+ "pressure": "pressure_down",
621
+ "pressure_arrow": "\u2198",
622
+ "pressure_display": "goal \u2193 | \u2198 -16.08% worsening",
623
+ "pressure_score": -1.0,
624
+ "trend_label": "worsening"
625
+ },
626
+ "weighted_regularizer_to_contrastive_ratio": {
627
+ "desired_direction": "lower",
628
+ "early_mean": 0.0667222202939108,
629
+ "goal_display": "goal \u2193",
630
+ "improvement_pct": 54.60656419810926,
631
+ "improvement_pct_display": "+54.61%",
632
+ "last": 0.028996589606316913,
633
+ "late_mean": 0.03028750823471251,
634
+ "mid_mean": 0.0329437448262181,
635
+ "pressure": "pressure_up",
636
+ "pressure_arrow": "\u2197",
637
+ "pressure_display": "goal \u2193 | \u2197 +54.61% improving",
638
+ "pressure_score": 1.0,
639
+ "trend_label": "improving"
640
+ }
641
+ },
642
+ "debug_metrics": {
643
+ "contrastive_loss": 0.5768923759460449,
644
+ "covariance_loss": 10.012618064880371,
645
+ "dataloader_wait_pct_300s": 0.0382592069912683,
646
+ "dataloader_wait_pct_60s": 0.0961140629935623,
647
+ "dataloader_wait_pct_900s": 0.066640439237091,
648
+ "dataloader_wait_pct_run": 0.06717802801040498,
649
+ "gpu_active_fraction": 1.0,
650
+ "gpu_max_memory_gb": 13.279157161712646,
651
+ "gpu_mem_used_mean_gb": 14.1943359375,
652
+ "gpu_mem_util_mean_pct": 14.848754175733243,
653
+ "gpu_snapshot_available": 1.0,
654
+ "gpu_util_max_pct": 54.0,
655
+ "gpu_util_mean_pct": 36.0,
656
+ "gpu_util_min_pct": 18.0,
657
+ "grad_norm": 1.7659728527069092,
658
+ "mfu_estimate": 0.06267708747552383,
659
+ "samples_per_sec": 1399.0203442084614,
660
+ "samples_per_sec_300s": 1547.033197547462,
661
+ "samples_per_sec_60s": 1399.0203442084614,
662
+ "samples_per_sec_900s": 1512.7885250414204,
663
+ "samples_per_sec_run_mean": 1533.0231795943853,
664
+ "step_time_cv_300s": 0.06743440395418321,
665
+ "step_time_cv_900s": 0.09218780075615579,
666
+ "step_time_sec": 0.3659703749981418,
667
+ "structure_bce_loss": 0.08383000642061234,
668
+ "structure_close_fraction": 0.029296875,
669
+ "structure_confidence_candidate_count": 2126.0,
670
+ "structure_confidence_separation": -0.0040241447277367115,
671
+ "structure_correct_logit_mean": 9.1875,
672
+ "structure_fingerprint_cosine": 0.46021854877471924,
673
+ "structure_hit_at_1": 0.00390625,
674
+ "structure_hit_at_20": 0.1171875,
675
+ "structure_hit_at_5": 0.029296875,
676
+ "structure_logit_margin_mean": -1.0078125,
677
+ "structure_retrieval_logit_std": 0.9839304089546204,
678
+ "structure_retrieval_loss": 10.630162239074707,
679
+ "structure_target_rank_mean": 314.703125,
680
+ "structure_top1_top2_gap_mean": 0.0986328125,
681
+ "structure_top1_top2_gap_norm_mean": 0.10155132412910461,
682
+ "structure_top_negative_logit_mean": 10.1875,
683
+ "structure_uncertain_fraction": 0.970703125,
684
+ "structure_uncertain_hit_at_1": 0.0040241447277367115,
685
+ "train_loss": 6.491608619689941,
686
+ "train_loss_ema": 6.519052214060582,
687
+ "val_loss": 6.646750174145564,
688
+ "val_ssl_loss": 1.0397093277424574,
689
+ "val_structure_close_fraction": 0.0077972412109375,
690
+ "val_structure_close_hit_at_1": 0.014835291571216658,
691
+ "val_structure_confidence_candidate_count": 4096.0,
692
+ "val_structure_confidence_separation": -0.027957430157584895,
693
+ "val_structure_correct_logit_mean": 9.49462890625,
694
+ "val_structure_retrieval_logit_std": 2.6622845763340592,
695
+ "val_structure_retrieval_loss": 10.882395941764116,
696
+ "val_structure_top1_top2_gap_mean": 0.11855697631835938,
697
+ "val_structure_top1_top2_gap_norm_mean": 0.05411380449368153,
698
+ "val_structure_top_negative_logit_mean": 10.714599609375,
699
+ "val_structure_uncertain_fraction": 0.9922027587890625,
700
+ "val_structure_uncertain_hit_at_1": 0.027957430157584895,
701
+ "variance_loss": 0.00213623046875,
702
+ "weighted_regularizer_to_contrastive_ratio": 0.028996589606316913
703
+ },
704
+ "eval_coverage": {
705
+ "configured_test_shard_count": 2,
706
+ "configured_test_shards": [
707
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051433_00000.parquet",
708
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051454_00000.parquet"
709
+ ],
710
+ "configured_val_shard_count": 2,
711
+ "configured_val_shards": [
712
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
713
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
714
+ ],
715
+ "created_at": "2026-04-13T21:36:17Z",
716
+ "eval_batch_limit": 256,
717
+ "eval_batches_per_shard_limit": 0,
718
+ "eval_loader_mode": "stream",
719
+ "eval_mode": "deep",
720
+ "full_holdout_coverage": 0.0,
721
+ "full_holdout_sweep": false,
722
+ "full_test_coverage": 0.0,
723
+ "full_val_coverage": 1.0,
724
+ "missing_test_shards": [
725
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051433_00000.parquet",
726
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051454_00000.parquet"
727
+ ],
728
+ "missing_val_shards": [],
729
+ "observed_eval_batches": 256,
730
+ "observed_eval_samples": 131072,
731
+ "observed_test_shard_count": 0,
732
+ "observed_test_shards": [],
733
+ "observed_val_shard_count": 2,
734
+ "observed_val_shards": [
735
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
736
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
737
+ ],
738
+ "per_shard_batch_count": {
739
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet": 128,
740
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet": 128
741
+ },
742
+ "per_shard_sample_count": {
743
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet": 65536,
744
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet": 65536
745
+ },
746
+ "run_step": 10000,
747
+ "step": 91001,
748
+ "unexpected_eval_shards": []
749
+ },
750
+ "failure_modes": [
751
+ {
752
+ "detail": "Observed shard counts {'train': 4, 'val': 2, 'test': 2}; the canonical Day 1 loader contract requires all splits to be readable.",
753
+ "name": "dataset_shards_accessible",
754
+ "status": "pass"
755
+ },
756
+ {
757
+ "detail": "Training batches continued to arrive with positive throughput and no iterator stall.",
758
+ "name": "dataloader_streaming_stable",
759
+ "status": "pass"
760
+ },
761
+ {
762
+ "detail": "Technical_Plan requires zero NaN or Inf tensors and finite train/val loss.",
763
+ "name": "non_finite_loss_or_tensor_failure",
764
+ "status": "pass"
765
+ },
766
+ {
767
+ "detail": "Engineering.md expects slope-aware monitoring; missing validation or retrieval metrics make trend judgment weaker.",
768
+ "name": "validation_observability_gap",
769
+ "status": "pass"
770
+ },
771
+ {
772
+ "detail": "Technical_Plan Day 1 success requires stable GPU memory through the bounded run.",
773
+ "name": "gpu_memory_instability",
774
+ "status": "pass"
775
+ },
776
+ {
777
+ "detail": "Checkpoint save and reload is an explicit Day 1 success criterion.",
778
+ "name": "checkpoint_save_reload_failure",
779
+ "status": "pass"
780
+ },
781
+ {
782
+ "detail": "Engineering.md expects embedding variance monitoring and non-collapsed representations.",
783
+ "name": "embedding_collapse",
784
+ "status": "pass"
785
+ },
786
+ {
787
+ "detail": "The current instability mode is regularization overpowering the SSL objective and producing jagged total loss.",
788
+ "name": "regularizer_dominance",
789
+ "status": "pass"
790
+ },
791
+ {
792
+ "detail": "Once the bridge phase begins, Technical_Plan requires retrieval-oriented validation rather than only augmented contrastive loss.",
793
+ "name": "structure_supervision_gap",
794
+ "status": "pass"
795
+ },
796
+ {
797
+ "detail": "Structure-first qualification now requires explicit confidence diagnostics such as top-1/top-2 separation and uncertainty buckets, not just broader retrieval metrics.",
798
+ "name": "confidence_surface_gap",
799
+ "status": "pass"
800
+ },
801
+ {
802
+ "detail": "The bridge phase should prefer chemistry-grounded fingerprint targets; hashed fallback targets weaken the retrieval signal.",
803
+ "name": "chemistry_target_backend_gap",
804
+ "status": "pass"
805
+ },
806
+ {
807
+ "detail": "Bridge runs must improve retrieval after fingerprint fit saturates; fitting BCE while val Hit@20 stays at candidate-bank baseline means ranking is not generalizing.",
808
+ "name": "bridge_fit_without_ranking",
809
+ "status": "pass"
810
+ },
811
+ {
812
+ "detail": "Engineering.md flags acquisition-artefact shortcut risk; adduct and instrument embeddings remain deliberately down-scaled.",
813
+ "name": "metadata_shortcutting",
814
+ "status": "watch"
815
+ }
816
+ ],
817
+ "fast_eval_surface": {
818
+ "batch_size": 512,
819
+ "batches_per_shard": 0,
820
+ "configured_val_shard_count": 2,
821
+ "configured_val_shards": [
822
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
823
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
824
+ ],
825
+ "created_at": "2026-04-13T20:20:58Z",
826
+ "loader_mode": "stream",
827
+ "per_shard_batch_count": {},
828
+ "per_shard_row_count": {},
829
+ "row_count": 0
830
+ },
831
+ "metric_pressure_summary": {
832
+ "average_improvement_pct": 34.52368967199551,
833
+ "metric_count": 39,
834
+ "overall_arrow": "\u2197",
835
+ "overall_pressure": "pressure_up",
836
+ "pressure_down_count": 11,
837
+ "pressure_none_count": 7,
838
+ "pressure_up_count": 21,
839
+ "pressure_up_fraction": 0.5384615384615384,
840
+ "weighted_improvement_pct": 21.88570985770741,
841
+ "weighted_pressure_score": 0.2838709677419355
842
+ },
843
+ "metric_pressure_summary_display": {
844
+ "average_improvement_pct": 34.52368967199551,
845
+ "average_improvement_pct_display": "+34.52%",
846
+ "metric_count": 39,
847
+ "overall_arrow": "\u2197",
848
+ "overall_display": "\u2197 improving (+21.89% weighted)",
849
+ "overall_label": "improving",
850
+ "overall_pressure": "pressure_up",
851
+ "pressure_down_count": 11,
852
+ "pressure_none_count": 7,
853
+ "pressure_up_count": 21,
854
+ "pressure_up_fraction": 0.5384615384615384,
855
+ "pressure_up_fraction_display": "+53.85%",
856
+ "weighted_improvement_pct": 21.88570985770741,
857
+ "weighted_improvement_pct_display": "+21.89%",
858
+ "weighted_pressure_score": 0.2838709677419355
859
+ },
860
+ "metric_pressures": {
861
+ "contrastive_loss": {
862
+ "desired_direction": "lower",
863
+ "early_mean": 1.409792113166324,
864
+ "improvement_pct": -1.6400841264380104,
865
+ "improvement_pct_clipped": -1.6400841264380104,
866
+ "last": 0.5768923759460449,
867
+ "late_mean": 1.43291388983014,
868
+ "mid_mean": 1.3598130204084982,
869
+ "pressure": "pressure_none",
870
+ "pressure_arrow": "\u2192",
871
+ "pressure_score": 0.0
872
+ },
873
+ "gpu_max_memory_gb": {
874
+ "desired_direction": "stable",
875
+ "early_mean": 13.252925169950275,
876
+ "improvement_pct": 100.0,
877
+ "improvement_pct_clipped": 100.0,
878
+ "last": 13.279157161712646,
879
+ "late_mean": 13.279157161712646,
880
+ "mid_mean": 13.279059082097401,
881
+ "pressure": "pressure_up",
882
+ "pressure_arrow": "\u2197",
883
+ "pressure_score": 1.0
884
+ },
885
+ "grad_norm": {
886
+ "desired_direction": "stable",
887
+ "early_mean": 4.638080189338309,
888
+ "improvement_pct": -100.0,
889
+ "improvement_pct_clipped": -100.0,
890
+ "last": 1.7659728527069092,
891
+ "late_mean": 6.072961742165445,
892
+ "mid_mean": 4.979553528603791,
893
+ "pressure": "pressure_down",
894
+ "pressure_arrow": "\u2198",
895
+ "pressure_score": -1.0
896
+ },
897
+ "samples_per_sec": {
898
+ "desired_direction": "higher",
899
+ "early_mean": 1575.5664566260843,
900
+ "improvement_pct": -2.8825714693532967,
901
+ "improvement_pct_clipped": -2.8825714693532967,
902
+ "last": 1399.0203442084614,
903
+ "late_mean": 1530.1496274666802,
904
+ "mid_mean": 1531.3735887168505,
905
+ "pressure": "pressure_down",
906
+ "pressure_arrow": "\u2198",
907
+ "pressure_score": -1.0
908
+ },
909
+ "structure_bce_loss": {
910
+ "desired_direction": "lower",
911
+ "early_mean": 0.12293951650489272,
912
+ "improvement_pct": 39.749156913940595,
913
+ "improvement_pct_clipped": 39.749156913940595,
914
+ "last": 0.08383000642061234,
915
+ "late_mean": 0.07407209518012302,
916
+ "mid_mean": 0.07652380578004556,
917
+ "pressure": "pressure_up",
918
+ "pressure_arrow": "\u2197",
919
+ "pressure_score": 1.0
920
+ },
921
+ "structure_confidence_separation": {
922
+ "desired_direction": "higher",
923
+ "early_mean": -0.007918493679557749,
924
+ "improvement_pct": -531.5617836245987,
925
+ "improvement_pct_clipped": -100.0,
926
+ "last": -0.0040241447277367115,
927
+ "late_mean": -0.05001017991881603,
928
+ "mid_mean": -0.032136323077151974,
929
+ "pressure": "pressure_down",
930
+ "pressure_arrow": "\u2198",
931
+ "pressure_score": -1.0
932
+ },
933
+ "structure_correct_logit_mean": {
934
+ "desired_direction": "higher",
935
+ "early_mean": 6.134031791907514,
936
+ "improvement_pct": 47.305071917778086,
937
+ "improvement_pct_clipped": 47.305071917778086,
938
+ "last": 9.1875,
939
+ "late_mean": 9.035739942528735,
940
+ "mid_mean": 9.474891618497109,
941
+ "pressure": "pressure_up",
942
+ "pressure_arrow": "\u2197",
943
+ "pressure_score": 1.0
944
+ },
945
+ "structure_fingerprint_cosine": {
946
+ "desired_direction": "higher",
947
+ "early_mean": 0.3965519580968543,
948
+ "improvement_pct": 17.58747890372286,
949
+ "improvement_pct_clipped": 17.58747890372286,
950
+ "last": 0.46021854877471924,
951
+ "late_mean": 0.46629545006943846,
952
+ "mid_mean": 0.4464927057310336,
953
+ "pressure": "pressure_up",
954
+ "pressure_arrow": "\u2197",
955
+ "pressure_score": 1.0
956
+ },
957
+ "structure_hit_at_1": {
958
+ "desired_direction": "higher",
959
+ "early_mean": 0.007880238439306358,
960
+ "improvement_pct": 549.6830023383724,
961
+ "improvement_pct_clipped": 100.0,
962
+ "last": 0.00390625,
963
+ "late_mean": 0.051196569683908046,
964
+ "mid_mean": 0.032706376445086706,
965
+ "pressure": "pressure_up",
966
+ "pressure_arrow": "\u2197",
967
+ "pressure_score": 1.0
968
+ },
969
+ "structure_hit_at_20": {
970
+ "desired_direction": "higher",
971
+ "early_mean": 0.10721865968208093,
972
+ "improvement_pct": 239.4614028144399,
973
+ "improvement_pct_clipped": 100.0,
974
+ "last": 0.1171875,
975
+ "late_mean": 0.3639659662356322,
976
+ "mid_mean": 0.28105807442196534,
977
+ "pressure": "pressure_up",
978
+ "pressure_arrow": "\u2197",
979
+ "pressure_score": 1.0
980
+ },
981
+ "structure_hit_at_5": {
982
+ "desired_direction": "higher",
983
+ "early_mean": 0.03360955563583815,
984
+ "improvement_pct": 402.47008675709174,
985
+ "improvement_pct_clipped": 100.0,
986
+ "last": 0.029296875,
987
+ "late_mean": 0.16887796336206898,
988
+ "mid_mean": 0.11673591040462428,
989
+ "pressure": "pressure_up",
990
+ "pressure_arrow": "\u2197",
991
+ "pressure_score": 1.0
992
+ },
993
+ "structure_logit_margin_mean": {
994
+ "desired_direction": "higher",
995
+ "early_mean": -1.1263999277456647,
996
+ "improvement_pct": 1.2274364862230593,
997
+ "improvement_pct_clipped": 1.2274364862230593,
998
+ "last": -1.0078125,
999
+ "late_mean": -1.1125740840517242,
1000
+ "mid_mean": -1.0154725885115607,
1001
+ "pressure": "pressure_none",
1002
+ "pressure_arrow": "\u2192",
1003
+ "pressure_score": 0.0
1004
+ },
1005
+ "structure_retrieval_logit_std": {
1006
+ "desired_direction": "higher",
1007
+ "early_mean": 1.0455416592559375,
1008
+ "improvement_pct": 112.05988926255111,
1009
+ "improvement_pct_clipped": 100.0,
1010
+ "last": 0.9839304089546204,
1011
+ "late_mean": 2.2171744848119803,
1012
+ "mid_mean": 2.0036371215919537,
1013
+ "pressure": "pressure_up",
1014
+ "pressure_arrow": "\u2197",
1015
+ "pressure_score": 1.0
1016
+ },
1017
+ "structure_retrieval_loss": {
1018
+ "desired_direction": "lower",
1019
+ "early_mean": 10.237479978903181,
1020
+ "improvement_pct": 10.094066619063353,
1021
+ "improvement_pct_clipped": 10.094066619063353,
1022
+ "last": 10.630162239074707,
1023
+ "late_mean": 9.204101929719421,
1024
+ "mid_mean": 9.41523380775672,
1025
+ "pressure": "pressure_up",
1026
+ "pressure_arrow": "\u2197",
1027
+ "pressure_score": 1.0
1028
+ },
1029
+ "structure_target_rank_mean": {
1030
+ "desired_direction": "lower",
1031
+ "early_mean": 351.84337743858384,
1032
+ "improvement_pct": 58.80171764880722,
1033
+ "improvement_pct_clipped": 58.80171764880722,
1034
+ "last": 314.703125,
1035
+ "late_mean": 144.9534280711207,
1036
+ "mid_mean": 176.16389315390174,
1037
+ "pressure": "pressure_up",
1038
+ "pressure_arrow": "\u2197",
1039
+ "pressure_score": 1.0
1040
+ },
1041
+ "structure_top1_top2_gap_mean": {
1042
+ "desired_direction": "higher",
1043
+ "early_mean": 0.09793567106213873,
1044
+ "improvement_pct": 59.221149173532794,
1045
+ "improvement_pct_clipped": 59.221149173532794,
1046
+ "last": 0.0986328125,
1047
+ "late_mean": 0.1559343009159483,
1048
+ "mid_mean": 0.12284365968208093,
1049
+ "pressure": "pressure_up",
1050
+ "pressure_arrow": "\u2197",
1051
+ "pressure_score": 1.0
1052
+ },
1053
+ "structure_top1_top2_gap_norm_mean": {
1054
+ "desired_direction": "higher",
1055
+ "early_mean": 0.12215662799301864,
1056
+ "improvement_pct": -35.653274040957065,
1057
+ "improvement_pct_clipped": -35.653274040957065,
1058
+ "last": 0.10155132412910461,
1059
+ "late_mean": 0.07860379065547524,
1060
+ "mid_mean": 0.07032911504236604,
1061
+ "pressure": "pressure_down",
1062
+ "pressure_arrow": "\u2198",
1063
+ "pressure_score": -1.0
1064
+ },
1065
+ "structure_top_negative_logit_mean": {
1066
+ "desired_direction": "lower",
1067
+ "early_mean": 7.262102601156069,
1068
+ "improvement_pct": -39.724147080005764,
1069
+ "improvement_pct_clipped": -39.724147080005764,
1070
+ "last": 10.1875,
1071
+ "late_mean": 10.14691091954023,
1072
+ "mid_mean": 10.490606936416185,
1073
+ "pressure": "pressure_down",
1074
+ "pressure_arrow": "\u2198",
1075
+ "pressure_score": -1.0
1076
+ },
1077
+ "structure_uncertain_fraction": {
1078
+ "desired_direction": "lower",
1079
+ "early_mean": 0.924663565751445,
1080
+ "improvement_pct": -6.108024336610137,
1081
+ "improvement_pct_clipped": -6.108024336610137,
1082
+ "last": 0.970703125,
1083
+ "late_mean": 0.9811422413793104,
1084
+ "mid_mean": 0.984747561416185,
1085
+ "pressure": "pressure_down",
1086
+ "pressure_arrow": "\u2198",
1087
+ "pressure_score": -1.0
1088
+ },
1089
+ "train_loss": {
1090
+ "desired_direction": "lower",
1091
+ "early_mean": 6.340343519442343,
1092
+ "improvement_pct": 10.566866163080938,
1093
+ "improvement_pct_clipped": 10.566866163080938,
1094
+ "last": 6.491608619689941,
1095
+ "late_mean": 5.670367905463295,
1096
+ "mid_mean": 5.795929590401622,
1097
+ "pressure": "pressure_up",
1098
+ "pressure_arrow": "\u2197",
1099
+ "pressure_score": 1.0
1100
+ },
1101
+ "train_loss_ema": {
1102
+ "desired_direction": "lower",
1103
+ "early_mean": 6.265581717362126,
1104
+ "improvement_pct": 10.455707205061195,
1105
+ "improvement_pct_clipped": 10.455707205061195,
1106
+ "last": 6.519052214060582,
1107
+ "late_mean": 5.610470838300897,
1108
+ "mid_mean": 5.8371378655061745,
1109
+ "pressure": "pressure_up",
1110
+ "pressure_arrow": "\u2197",
1111
+ "pressure_score": 1.0
1112
+ },
1113
+ "val_hit_at_1": {
1114
+ "desired_direction": "higher",
1115
+ "early_mean": 0.0079345703125,
1116
+ "improvement_pct": 207.78846153846155,
1117
+ "improvement_pct_clipped": 100.0,
1118
+ "last": 0.0282135009765625,
1119
+ "late_mean": 0.02442169189453125,
1120
+ "mid_mean": 0.01372528076171875,
1121
+ "pressure": "pressure_none",
1122
+ "pressure_arrow": "\u2192",
1123
+ "pressure_score": 0.0
1124
+ },
1125
+ "val_hit_at_20": {
1126
+ "desired_direction": "higher",
1127
+ "early_mean": 0.08158111572265625,
1128
+ "improvement_pct": 76.67165435331526,
1129
+ "improvement_pct_clipped": 76.67165435331526,
1130
+ "last": 0.15270233154296875,
1131
+ "late_mean": 0.14413070678710938,
1132
+ "mid_mean": 0.094146728515625,
1133
+ "pressure": "pressure_up",
1134
+ "pressure_arrow": "\u2197",
1135
+ "pressure_score": 1.0
1136
+ },
1137
+ "val_hit_at_5": {
1138
+ "desired_direction": "higher",
1139
+ "early_mean": 0.0288543701171875,
1140
+ "improvement_pct": 134.8492860920148,
1141
+ "improvement_pct_clipped": 100.0,
1142
+ "last": 0.0726776123046875,
1143
+ "late_mean": 0.0677642822265625,
1144
+ "mid_mean": 0.0425567626953125,
1145
+ "pressure": "pressure_up",
1146
+ "pressure_arrow": "\u2197",
1147
+ "pressure_score": 1.0
1148
+ },
1149
+ "val_loss": {
1150
+ "desired_direction": "lower",
1151
+ "early_mean": 1.2876316868363775,
1152
+ "improvement_pct": -55.82459156154955,
1153
+ "improvement_pct_clipped": -55.82459156154955,
1154
+ "last": 6.646750174145564,
1155
+ "late_mean": 2.006446816829876,
1156
+ "mid_mean": 2.20697861268016,
1157
+ "pressure": "pressure_down",
1158
+ "pressure_arrow": "\u2198",
1159
+ "pressure_score": -1.0
1160
+ },
1161
+ "val_ssl_loss": {
1162
+ "desired_direction": "lower",
1163
+ "early_mean": 1.0849772497183745,
1164
+ "improvement_pct": 2.3192742421198114,
1165
+ "improvement_pct_clipped": 2.3192742421198114,
1166
+ "last": 1.0397093277424574,
1167
+ "late_mean": 1.0598136518327963,
1168
+ "mid_mean": 1.0747967428483425,
1169
+ "pressure": "pressure_none",
1170
+ "pressure_arrow": "\u2192",
1171
+ "pressure_score": 0.0
1172
+ },
1173
+ "val_structure_bce_loss": {
1174
+ "desired_direction": "lower",
1175
+ "early_mean": 0.0771060927848642,
1176
+ "improvement_pct": 12.199713989530546,
1177
+ "improvement_pct_clipped": 12.199713989530546,
1178
+ "last": 0.06532714269997086,
1179
+ "late_mean": 0.06769936999660872,
1180
+ "mid_mean": 0.07029345531468391,
1181
+ "pressure": "pressure_up",
1182
+ "pressure_arrow": "\u2197",
1183
+ "pressure_score": 1.0
1184
+ },
1185
+ "val_structure_confidence_separation": {
1186
+ "desired_direction": "higher",
1187
+ "early_mean": -0.00795668972114072,
1188
+ "improvement_pct": -204.85499246505236,
1189
+ "improvement_pct_clipped": -100.0,
1190
+ "last": -0.027957430157584895,
1191
+ "late_mean": -0.024256365849851136,
1192
+ "mid_mean": -0.01309866098972634,
1193
+ "pressure": "pressure_down",
1194
+ "pressure_arrow": "\u2198",
1195
+ "pressure_score": -1.0
1196
+ },
1197
+ "val_structure_correct_logit_mean": {
1198
+ "desired_direction": "higher",
1199
+ "early_mean": 8.0311279296875,
1200
+ "improvement_pct": 16.44145855816145,
1201
+ "improvement_pct_clipped": 16.44145855816145,
1202
+ "last": 9.49462890625,
1203
+ "late_mean": 9.3515625,
1204
+ "mid_mean": 8.441162109375,
1205
+ "pressure": "pressure_up",
1206
+ "pressure_arrow": "\u2197",
1207
+ "pressure_score": 1.0
1208
+ },
1209
+ "val_structure_fingerprint_cosine": {
1210
+ "desired_direction": "higher",
1211
+ "early_mean": 0.4147872725734487,
1212
+ "improvement_pct": 0.17175282510408285,
1213
+ "improvement_pct_clipped": 0.17175282510408285,
1214
+ "last": 0.4255251376889646,
1215
+ "late_mean": 0.4154996814322658,
1216
+ "mid_mean": 0.4204998221830465,
1217
+ "pressure": "pressure_none",
1218
+ "pressure_arrow": "\u2192",
1219
+ "pressure_score": 0.0
1220
+ },
1221
+ "val_structure_logit_margin_mean": {
1222
+ "desired_direction": "higher",
1223
+ "early_mean": -1.194091796875,
1224
+ "improvement_pct": -13.669239419341647,
1225
+ "improvement_pct_clipped": -13.669239419341647,
1226
+ "last": -1.221923828125,
1227
+ "late_mean": -1.3573150634765625,
1228
+ "mid_mean": -2.0707931518554688,
1229
+ "pressure": "pressure_down",
1230
+ "pressure_arrow": "\u2198",
1231
+ "pressure_score": -1.0
1232
+ },
1233
+ "val_structure_retrieval_logit_std": {
1234
+ "desired_direction": "higher",
1235
+ "early_mean": 1.2326001720502973,
1236
+ "improvement_pct": 102.32641409365677,
1237
+ "improvement_pct_clipped": 100.0,
1238
+ "last": 2.6622845763340592,
1239
+ "late_mean": 2.49387572822161,
1240
+ "mid_mean": 2.022283688886091,
1241
+ "pressure": "pressure_up",
1242
+ "pressure_arrow": "\u2197",
1243
+ "pressure_score": 1.0
1244
+ },
1245
+ "val_structure_retrieval_loss": {
1246
+ "desired_direction": "lower",
1247
+ "early_mean": 11.562767315655947,
1248
+ "improvement_pct": 4.864921250164783,
1249
+ "improvement_pct_clipped": 4.864921250164783,
1250
+ "last": 10.882395941764116,
1251
+ "late_mean": 11.000247791409492,
1252
+ "mid_mean": 12.150514367967844,
1253
+ "pressure": "pressure_none",
1254
+ "pressure_arrow": "\u2192",
1255
+ "pressure_score": 0.0
1256
+ },
1257
+ "val_structure_target_rank_mean": {
1258
+ "desired_direction": "lower",
1259
+ "early_mean": 760.7540435791016,
1260
+ "improvement_pct": 45.44775026271755,
1261
+ "improvement_pct_clipped": 45.44775026271755,
1262
+ "last": 379.8781280517578,
1263
+ "late_mean": 415.0084457397461,
1264
+ "mid_mean": 683.3385162353516,
1265
+ "pressure": "pressure_up",
1266
+ "pressure_arrow": "\u2197",
1267
+ "pressure_score": 1.0
1268
+ },
1269
+ "val_structure_top1_top2_gap_mean": {
1270
+ "desired_direction": "higher",
1271
+ "early_mean": 0.07839679718017578,
1272
+ "improvement_pct": 54.90785232041847,
1273
+ "improvement_pct_clipped": 54.90785232041847,
1274
+ "last": 0.11855697631835938,
1275
+ "late_mean": 0.12144279479980469,
1276
+ "mid_mean": 0.1310901641845703,
1277
+ "pressure": "pressure_up",
1278
+ "pressure_arrow": "\u2197",
1279
+ "pressure_score": 1.0
1280
+ },
1281
+ "val_structure_top1_top2_gap_norm_mean": {
1282
+ "desired_direction": "higher",
1283
+ "early_mean": 0.06950851557485294,
1284
+ "improvement_pct": -16.826381575968767,
1285
+ "improvement_pct_clipped": -16.826381575968767,
1286
+ "last": 0.05411380449368153,
1287
+ "late_mean": 0.0578127475164365,
1288
+ "mid_mean": 0.07577372609375743,
1289
+ "pressure": "pressure_down",
1290
+ "pressure_arrow": "\u2198",
1291
+ "pressure_score": -1.0
1292
+ },
1293
+ "val_structure_top_negative_logit_mean": {
1294
+ "desired_direction": "lower",
1295
+ "early_mean": 9.2255859375,
1296
+ "improvement_pct": -16.07653223245475,
1297
+ "improvement_pct_clipped": -16.07653223245475,
1298
+ "last": 10.714599609375,
1299
+ "late_mean": 10.708740234375,
1300
+ "mid_mean": 10.51171875,
1301
+ "pressure": "pressure_down",
1302
+ "pressure_arrow": "\u2198",
1303
+ "pressure_score": -1.0
1304
+ },
1305
+ "val_structure_uncertain_fraction": {
1306
+ "desired_direction": "lower",
1307
+ "early_mean": 0.9941177368164062,
1308
+ "improvement_pct": -0.03261678728482514,
1309
+ "improvement_pct_clipped": -0.03261678728482514,
1310
+ "last": 0.9922027587890625,
1311
+ "late_mean": 0.9944419860839844,
1312
+ "mid_mean": 0.9762382507324219,
1313
+ "pressure": "pressure_none",
1314
+ "pressure_arrow": "\u2192",
1315
+ "pressure_score": 0.0
1316
+ },
1317
+ "weighted_regularizer_to_contrastive_ratio": {
1318
+ "desired_direction": "lower",
1319
+ "early_mean": 0.0667222202939108,
1320
+ "improvement_pct": 54.60656419810926,
1321
+ "improvement_pct_clipped": 54.60656419810926,
1322
+ "last": 0.028996589606316913,
1323
+ "late_mean": 0.03028750823471251,
1324
+ "mid_mean": 0.0329437448262181,
1325
+ "pressure": "pressure_up",
1326
+ "pressure_arrow": "\u2197",
1327
+ "pressure_score": 1.0
1328
+ }
1329
+ },
1330
+ "metric_pressures_display": {
1331
+ "contrastive_loss": {
1332
+ "desired_direction": "lower",
1333
+ "early_mean": 1.409792113166324,
1334
+ "goal_display": "goal \u2193",
1335
+ "improvement_pct": -1.6400841264380104,
1336
+ "improvement_pct_display": "-1.64%",
1337
+ "last": 0.5768923759460449,
1338
+ "late_mean": 1.43291388983014,
1339
+ "mid_mean": 1.3598130204084982,
1340
+ "pressure": "pressure_none",
1341
+ "pressure_arrow": "\u2192",
1342
+ "pressure_display": "goal \u2193 | \u2192 -1.64% flat",
1343
+ "trend_label": "flat"
1344
+ },
1345
+ "gpu_max_memory_gb": {
1346
+ "desired_direction": "stable",
1347
+ "early_mean": 13.252925169950275,
1348
+ "goal_display": "goal \u2248",
1349
+ "improvement_pct": 100.0,
1350
+ "improvement_pct_display": "+100.00%",
1351
+ "last": 13.279157161712646,
1352
+ "late_mean": 13.279157161712646,
1353
+ "mid_mean": 13.279059082097401,
1354
+ "pressure": "pressure_up",
1355
+ "pressure_arrow": "\u2197",
1356
+ "pressure_display": "goal \u2248 | \u2197 +100.00% improving",
1357
+ "pressure_score": 1.0,
1358
+ "trend_label": "improving"
1359
+ },
1360
+ "grad_norm": {
1361
+ "desired_direction": "stable",
1362
+ "early_mean": 4.638080189338309,
1363
+ "goal_display": "goal \u2248",
1364
+ "improvement_pct": -100.0,
1365
+ "improvement_pct_display": "-100.00%",
1366
+ "last": 1.7659728527069092,
1367
+ "late_mean": 6.072961742165445,
1368
+ "mid_mean": 4.979553528603791,
1369
+ "pressure": "pressure_down",
1370
+ "pressure_arrow": "\u2198",
1371
+ "pressure_display": "goal \u2248 | \u2198 -100.00% worsening",
1372
+ "pressure_score": -1.0,
1373
+ "trend_label": "worsening"
1374
+ },
1375
+ "samples_per_sec": {
1376
+ "desired_direction": "higher",
1377
+ "early_mean": 1575.5664566260843,
1378
+ "goal_display": "goal \u2191",
1379
+ "improvement_pct": -2.8825714693532967,
1380
+ "improvement_pct_display": "-2.88%",
1381
+ "last": 1399.0203442084614,
1382
+ "late_mean": 1530.1496274666802,
1383
+ "mid_mean": 1531.3735887168505,
1384
+ "pressure": "pressure_down",
1385
+ "pressure_arrow": "\u2198",
1386
+ "pressure_display": "goal \u2191 | \u2198 -2.88% worsening",
1387
+ "pressure_score": -1.0,
1388
+ "trend_label": "worsening"
1389
+ },
1390
+ "structure_bce_loss": {
1391
+ "desired_direction": "lower",
1392
+ "early_mean": 0.12293951650489272,
1393
+ "goal_display": "goal \u2193",
1394
+ "improvement_pct": 39.749156913940595,
1395
+ "improvement_pct_display": "+39.75%",
1396
+ "last": 0.08383000642061234,
1397
+ "late_mean": 0.07407209518012302,
1398
+ "mid_mean": 0.07652380578004556,
1399
+ "pressure": "pressure_up",
1400
+ "pressure_arrow": "\u2197",
1401
+ "pressure_display": "goal \u2193 | \u2197 +39.75% improving",
1402
+ "pressure_score": 1.0,
1403
+ "trend_label": "improving"
1404
+ },
1405
+ "structure_confidence_separation": {
1406
+ "desired_direction": "higher",
1407
+ "early_mean": -0.007918493679557749,
1408
+ "goal_display": "goal \u2191",
1409
+ "improvement_pct": -531.5617836245987,
1410
+ "improvement_pct_clipped": -100.0,
1411
+ "improvement_pct_clipped_display": "-100.00%",
1412
+ "improvement_pct_display": "-531.56%",
1413
+ "last": -0.0040241447277367115,
1414
+ "late_mean": -0.05001017991881603,
1415
+ "mid_mean": -0.032136323077151974,
1416
+ "pressure": "pressure_down",
1417
+ "pressure_arrow": "\u2198",
1418
+ "pressure_display": "goal \u2191 | \u2198 -531.56% worsening",
1419
+ "pressure_score": -1.0,
1420
+ "trend_label": "worsening"
1421
+ },
1422
+ "structure_correct_logit_mean": {
1423
+ "desired_direction": "higher",
1424
+ "early_mean": 6.134031791907514,
1425
+ "goal_display": "goal \u2191",
1426
+ "improvement_pct": 47.305071917778086,
1427
+ "improvement_pct_display": "+47.31%",
1428
+ "last": 9.1875,
1429
+ "late_mean": 9.035739942528735,
1430
+ "mid_mean": 9.474891618497109,
1431
+ "pressure": "pressure_up",
1432
+ "pressure_arrow": "\u2197",
1433
+ "pressure_display": "goal \u2191 | \u2197 +47.31% improving",
1434
+ "pressure_score": 1.0,
1435
+ "trend_label": "improving"
1436
+ },
1437
+ "structure_fingerprint_cosine": {
1438
+ "desired_direction": "higher",
1439
+ "early_mean": 0.3965519580968543,
1440
+ "goal_display": "goal \u2191",
1441
+ "improvement_pct": 17.58747890372286,
1442
+ "improvement_pct_display": "+17.59%",
1443
+ "last": 0.46021854877471924,
1444
+ "late_mean": 0.46629545006943846,
1445
+ "mid_mean": 0.4464927057310336,
1446
+ "pressure": "pressure_up",
1447
+ "pressure_arrow": "\u2197",
1448
+ "pressure_display": "goal \u2191 | \u2197 +17.59% improving",
1449
+ "pressure_score": 1.0,
1450
+ "trend_label": "improving"
1451
+ },
1452
+ "structure_hit_at_1": {
1453
+ "desired_direction": "higher",
1454
+ "early_mean": 0.007880238439306358,
1455
+ "goal_display": "goal \u2191",
1456
+ "improvement_pct": 549.6830023383724,
1457
+ "improvement_pct_clipped": 100.0,
1458
+ "improvement_pct_clipped_display": "+100.00%",
1459
+ "improvement_pct_display": "+549.68%",
1460
+ "last": 0.00390625,
1461
+ "late_mean": 0.051196569683908046,
1462
+ "mid_mean": 0.032706376445086706,
1463
+ "pressure": "pressure_up",
1464
+ "pressure_arrow": "\u2197",
1465
+ "pressure_display": "goal \u2191 | \u2197 +549.68% improving",
1466
+ "pressure_score": 1.0,
1467
+ "trend_label": "improving"
1468
+ },
1469
+ "structure_hit_at_20": {
1470
+ "desired_direction": "higher",
1471
+ "early_mean": 0.10721865968208093,
1472
+ "goal_display": "goal \u2191",
1473
+ "improvement_pct": 239.4614028144399,
1474
+ "improvement_pct_clipped": 100.0,
1475
+ "improvement_pct_clipped_display": "+100.00%",
1476
+ "improvement_pct_display": "+239.46%",
1477
+ "last": 0.1171875,
1478
+ "late_mean": 0.3639659662356322,
1479
+ "mid_mean": 0.28105807442196534,
1480
+ "pressure": "pressure_up",
1481
+ "pressure_arrow": "\u2197",
1482
+ "pressure_display": "goal \u2191 | \u2197 +239.46% improving",
1483
+ "pressure_score": 1.0,
1484
+ "trend_label": "improving"
1485
+ },
1486
+ "structure_hit_at_5": {
1487
+ "desired_direction": "higher",
1488
+ "early_mean": 0.03360955563583815,
1489
+ "goal_display": "goal \u2191",
1490
+ "improvement_pct": 402.47008675709174,
1491
+ "improvement_pct_clipped": 100.0,
1492
+ "improvement_pct_clipped_display": "+100.00%",
1493
+ "improvement_pct_display": "+402.47%",
1494
+ "last": 0.029296875,
1495
+ "late_mean": 0.16887796336206898,
1496
+ "mid_mean": 0.11673591040462428,
1497
+ "pressure": "pressure_up",
1498
+ "pressure_arrow": "\u2197",
1499
+ "pressure_display": "goal \u2191 | \u2197 +402.47% improving",
1500
+ "pressure_score": 1.0,
1501
+ "trend_label": "improving"
1502
+ },
1503
+ "structure_logit_margin_mean": {
1504
+ "desired_direction": "higher",
1505
+ "early_mean": -1.1263999277456647,
1506
+ "goal_display": "goal \u2191",
1507
+ "improvement_pct": 1.2274364862230593,
1508
+ "improvement_pct_display": "+1.23%",
1509
+ "last": -1.0078125,
1510
+ "late_mean": -1.1125740840517242,
1511
+ "mid_mean": -1.0154725885115607,
1512
+ "pressure": "pressure_none",
1513
+ "pressure_arrow": "\u2192",
1514
+ "pressure_display": "goal \u2191 | \u2192 +1.23% flat",
1515
+ "trend_label": "flat"
1516
+ },
1517
+ "structure_retrieval_logit_std": {
1518
+ "desired_direction": "higher",
1519
+ "early_mean": 1.0455416592559375,
1520
+ "goal_display": "goal \u2191",
1521
+ "improvement_pct": 112.05988926255111,
1522
+ "improvement_pct_clipped": 100.0,
1523
+ "improvement_pct_clipped_display": "+100.00%",
1524
+ "improvement_pct_display": "+112.06%",
1525
+ "last": 0.9839304089546204,
1526
+ "late_mean": 2.2171744848119803,
1527
+ "mid_mean": 2.0036371215919537,
1528
+ "pressure": "pressure_up",
1529
+ "pressure_arrow": "\u2197",
1530
+ "pressure_display": "goal \u2191 | \u2197 +112.06% improving",
1531
+ "pressure_score": 1.0,
1532
+ "trend_label": "improving"
1533
+ },
1534
+ "structure_retrieval_loss": {
1535
+ "desired_direction": "lower",
1536
+ "early_mean": 10.237479978903181,
1537
+ "goal_display": "goal \u2193",
1538
+ "improvement_pct": 10.094066619063353,
1539
+ "improvement_pct_display": "+10.09%",
1540
+ "last": 10.630162239074707,
1541
+ "late_mean": 9.204101929719421,
1542
+ "mid_mean": 9.41523380775672,
1543
+ "pressure": "pressure_up",
1544
+ "pressure_arrow": "\u2197",
1545
+ "pressure_display": "goal \u2193 | \u2197 +10.09% improving",
1546
+ "pressure_score": 1.0,
1547
+ "trend_label": "improving"
1548
+ },
1549
+ "structure_target_rank_mean": {
1550
+ "desired_direction": "lower",
1551
+ "early_mean": 351.84337743858384,
1552
+ "goal_display": "goal \u2193",
1553
+ "improvement_pct": 58.80171764880722,
1554
+ "improvement_pct_display": "+58.80%",
1555
+ "last": 314.703125,
1556
+ "late_mean": 144.9534280711207,
1557
+ "mid_mean": 176.16389315390174,
1558
+ "pressure": "pressure_up",
1559
+ "pressure_arrow": "\u2197",
1560
+ "pressure_display": "goal \u2193 | \u2197 +58.80% improving",
1561
+ "pressure_score": 1.0,
1562
+ "trend_label": "improving"
1563
+ },
1564
+ "structure_top1_top2_gap_mean": {
1565
+ "desired_direction": "higher",
1566
+ "early_mean": 0.09793567106213873,
1567
+ "goal_display": "goal \u2191",
1568
+ "improvement_pct": 59.221149173532794,
1569
+ "improvement_pct_display": "+59.22%",
1570
+ "last": 0.0986328125,
1571
+ "late_mean": 0.1559343009159483,
1572
+ "mid_mean": 0.12284365968208093,
1573
+ "pressure": "pressure_up",
1574
+ "pressure_arrow": "\u2197",
1575
+ "pressure_display": "goal \u2191 | \u2197 +59.22% improving",
1576
+ "pressure_score": 1.0,
1577
+ "trend_label": "improving"
1578
+ },
1579
+ "structure_top1_top2_gap_norm_mean": {
1580
+ "desired_direction": "higher",
1581
+ "early_mean": 0.12215662799301864,
1582
+ "goal_display": "goal \u2191",
1583
+ "improvement_pct": -35.653274040957065,
1584
+ "improvement_pct_display": "-35.65%",
1585
+ "last": 0.10155132412910461,
1586
+ "late_mean": 0.07860379065547524,
1587
+ "mid_mean": 0.07032911504236604,
1588
+ "pressure": "pressure_down",
1589
+ "pressure_arrow": "\u2198",
1590
+ "pressure_display": "goal \u2191 | \u2198 -35.65% worsening",
1591
+ "pressure_score": -1.0,
1592
+ "trend_label": "worsening"
1593
+ },
1594
+ "structure_top_negative_logit_mean": {
1595
+ "desired_direction": "lower",
1596
+ "early_mean": 7.262102601156069,
1597
+ "goal_display": "goal \u2193",
1598
+ "improvement_pct": -39.724147080005764,
1599
+ "improvement_pct_display": "-39.72%",
1600
+ "last": 10.1875,
1601
+ "late_mean": 10.14691091954023,
1602
+ "mid_mean": 10.490606936416185,
1603
+ "pressure": "pressure_down",
1604
+ "pressure_arrow": "\u2198",
1605
+ "pressure_display": "goal \u2193 | \u2198 -39.72% worsening",
1606
+ "pressure_score": -1.0,
1607
+ "trend_label": "worsening"
1608
+ },
1609
+ "structure_uncertain_fraction": {
1610
+ "desired_direction": "lower",
1611
+ "early_mean": 0.924663565751445,
1612
+ "goal_display": "goal \u2193",
1613
+ "improvement_pct": -6.108024336610137,
1614
+ "improvement_pct_display": "-6.11%",
1615
+ "last": 0.970703125,
1616
+ "late_mean": 0.9811422413793104,
1617
+ "mid_mean": 0.984747561416185,
1618
+ "pressure": "pressure_down",
1619
+ "pressure_arrow": "\u2198",
1620
+ "pressure_display": "goal \u2193 | \u2198 -6.11% worsening",
1621
+ "pressure_score": -1.0,
1622
+ "trend_label": "worsening"
1623
+ },
1624
+ "train_loss": {
1625
+ "desired_direction": "lower",
1626
+ "early_mean": 6.340343519442343,
1627
+ "goal_display": "goal \u2193",
1628
+ "improvement_pct": 10.566866163080938,
1629
+ "improvement_pct_display": "+10.57%",
1630
+ "last": 6.491608619689941,
1631
+ "late_mean": 5.670367905463295,
1632
+ "mid_mean": 5.795929590401622,
1633
+ "pressure": "pressure_up",
1634
+ "pressure_arrow": "\u2197",
1635
+ "pressure_display": "goal \u2193 | \u2197 +10.57% improving",
1636
+ "pressure_score": 1.0,
1637
+ "trend_label": "improving"
1638
+ },
1639
+ "train_loss_ema": {
1640
+ "desired_direction": "lower",
1641
+ "early_mean": 6.265581717362126,
1642
+ "goal_display": "goal \u2193",
1643
+ "improvement_pct": 10.455707205061195,
1644
+ "improvement_pct_display": "+10.46%",
1645
+ "last": 6.519052214060582,
1646
+ "late_mean": 5.610470838300897,
1647
+ "mid_mean": 5.8371378655061745,
1648
+ "pressure": "pressure_up",
1649
+ "pressure_arrow": "\u2197",
1650
+ "pressure_display": "goal \u2193 | \u2197 +10.46% improving",
1651
+ "pressure_score": 1.0,
1652
+ "trend_label": "improving"
1653
+ },
1654
+ "val_hit_at_1": {
1655
+ "desired_direction": "higher",
1656
+ "early_mean": 0.0079345703125,
1657
+ "goal_display": "goal \u2191",
1658
+ "improvement_pct": 207.78846153846155,
1659
+ "improvement_pct_clipped": 100.0,
1660
+ "improvement_pct_clipped_display": "+100.00%",
1661
+ "improvement_pct_display": "+207.79%",
1662
+ "last": 0.0282135009765625,
1663
+ "late_mean": 0.02442169189453125,
1664
+ "mid_mean": 0.01372528076171875,
1665
+ "pressure": "pressure_none",
1666
+ "pressure_arrow": "\u2192",
1667
+ "pressure_display": "goal \u2191 | \u2192 +207.79% flat",
1668
+ "trend_label": "flat"
1669
+ },
1670
+ "val_hit_at_20": {
1671
+ "desired_direction": "higher",
1672
+ "early_mean": 0.08158111572265625,
1673
+ "goal_display": "goal \u2191",
1674
+ "improvement_pct": 76.67165435331526,
1675
+ "improvement_pct_display": "+76.67%",
1676
+ "last": 0.15270233154296875,
1677
+ "late_mean": 0.14413070678710938,
1678
+ "mid_mean": 0.094146728515625,
1679
+ "pressure": "pressure_up",
1680
+ "pressure_arrow": "\u2197",
1681
+ "pressure_display": "goal \u2191 | \u2197 +76.67% improving",
1682
+ "pressure_score": 1.0,
1683
+ "trend_label": "improving"
1684
+ },
1685
+ "val_hit_at_5": {
1686
+ "desired_direction": "higher",
1687
+ "early_mean": 0.0288543701171875,
1688
+ "goal_display": "goal \u2191",
1689
+ "improvement_pct": 134.8492860920148,
1690
+ "improvement_pct_clipped": 100.0,
1691
+ "improvement_pct_clipped_display": "+100.00%",
1692
+ "improvement_pct_display": "+134.85%",
1693
+ "last": 0.0726776123046875,
1694
+ "late_mean": 0.0677642822265625,
1695
+ "mid_mean": 0.0425567626953125,
1696
+ "pressure": "pressure_up",
1697
+ "pressure_arrow": "\u2197",
1698
+ "pressure_display": "goal \u2191 | \u2197 +134.85% improving",
1699
+ "pressure_score": 1.0,
1700
+ "trend_label": "improving"
1701
+ },
1702
+ "val_loss": {
1703
+ "desired_direction": "lower",
1704
+ "early_mean": 1.2876316868363775,
1705
+ "goal_display": "goal \u2193",
1706
+ "improvement_pct": -55.82459156154955,
1707
+ "improvement_pct_display": "-55.82%",
1708
+ "last": 6.646750174145564,
1709
+ "late_mean": 2.006446816829876,
1710
+ "mid_mean": 2.20697861268016,
1711
+ "pressure": "pressure_down",
1712
+ "pressure_arrow": "\u2198",
1713
+ "pressure_display": "goal \u2193 | \u2198 -55.82% worsening",
1714
+ "pressure_score": -1.0,
1715
+ "trend_label": "worsening"
1716
+ },
1717
+ "val_ssl_loss": {
1718
+ "desired_direction": "lower",
1719
+ "early_mean": 1.0849772497183745,
1720
+ "goal_display": "goal \u2193",
1721
+ "improvement_pct": 2.3192742421198114,
1722
+ "improvement_pct_display": "+2.32%",
1723
+ "last": 1.0397093277424574,
1724
+ "late_mean": 1.0598136518327963,
1725
+ "mid_mean": 1.0747967428483425,
1726
+ "pressure": "pressure_none",
1727
+ "pressure_arrow": "\u2192",
1728
+ "pressure_display": "goal \u2193 | \u2192 +2.32% flat",
1729
+ "trend_label": "flat"
1730
+ },
1731
+ "val_structure_bce_loss": {
1732
+ "desired_direction": "lower",
1733
+ "early_mean": 0.0771060927848642,
1734
+ "goal_display": "goal \u2193",
1735
+ "improvement_pct": 12.199713989530546,
1736
+ "improvement_pct_display": "+12.20%",
1737
+ "last": 0.06532714269997086,
1738
+ "late_mean": 0.06769936999660872,
1739
+ "mid_mean": 0.07029345531468391,
1740
+ "pressure": "pressure_up",
1741
+ "pressure_arrow": "\u2197",
1742
+ "pressure_display": "goal \u2193 | \u2197 +12.20% improving",
1743
+ "pressure_score": 1.0,
1744
+ "trend_label": "improving"
1745
+ },
1746
+ "val_structure_confidence_separation": {
1747
+ "desired_direction": "higher",
1748
+ "early_mean": -0.00795668972114072,
1749
+ "goal_display": "goal \u2191",
1750
+ "improvement_pct": -204.85499246505236,
1751
+ "improvement_pct_clipped": -100.0,
1752
+ "improvement_pct_clipped_display": "-100.00%",
1753
+ "improvement_pct_display": "-204.85%",
1754
+ "last": -0.027957430157584895,
1755
+ "late_mean": -0.024256365849851136,
1756
+ "mid_mean": -0.01309866098972634,
1757
+ "pressure": "pressure_down",
1758
+ "pressure_arrow": "\u2198",
1759
+ "pressure_display": "goal \u2191 | \u2198 -204.85% worsening",
1760
+ "pressure_score": -1.0,
1761
+ "trend_label": "worsening"
1762
+ },
1763
+ "val_structure_correct_logit_mean": {
1764
+ "desired_direction": "higher",
1765
+ "early_mean": 8.0311279296875,
1766
+ "goal_display": "goal \u2191",
1767
+ "improvement_pct": 16.44145855816145,
1768
+ "improvement_pct_display": "+16.44%",
1769
+ "last": 9.49462890625,
1770
+ "late_mean": 9.3515625,
1771
+ "mid_mean": 8.441162109375,
1772
+ "pressure": "pressure_up",
1773
+ "pressure_arrow": "\u2197",
1774
+ "pressure_display": "goal \u2191 | \u2197 +16.44% improving",
1775
+ "pressure_score": 1.0,
1776
+ "trend_label": "improving"
1777
+ },
1778
+ "val_structure_fingerprint_cosine": {
1779
+ "desired_direction": "higher",
1780
+ "early_mean": 0.4147872725734487,
1781
+ "goal_display": "goal \u2191",
1782
+ "improvement_pct": 0.17175282510408285,
1783
+ "improvement_pct_display": "+0.17%",
1784
+ "last": 0.4255251376889646,
1785
+ "late_mean": 0.4154996814322658,
1786
+ "mid_mean": 0.4204998221830465,
1787
+ "pressure": "pressure_none",
1788
+ "pressure_arrow": "\u2192",
1789
+ "pressure_display": "goal \u2191 | \u2192 +0.17% flat",
1790
+ "trend_label": "flat"
1791
+ },
1792
+ "val_structure_logit_margin_mean": {
1793
+ "desired_direction": "higher",
1794
+ "early_mean": -1.194091796875,
1795
+ "goal_display": "goal \u2191",
1796
+ "improvement_pct": -13.669239419341647,
1797
+ "improvement_pct_display": "-13.67%",
1798
+ "last": -1.221923828125,
1799
+ "late_mean": -1.3573150634765625,
1800
+ "mid_mean": -2.0707931518554688,
1801
+ "pressure": "pressure_down",
1802
+ "pressure_arrow": "\u2198",
1803
+ "pressure_display": "goal \u2191 | \u2198 -13.67% worsening",
1804
+ "pressure_score": -1.0,
1805
+ "trend_label": "worsening"
1806
+ },
1807
+ "val_structure_retrieval_logit_std": {
1808
+ "desired_direction": "higher",
1809
+ "early_mean": 1.2326001720502973,
1810
+ "goal_display": "goal \u2191",
1811
+ "improvement_pct": 102.32641409365677,
1812
+ "improvement_pct_clipped": 100.0,
1813
+ "improvement_pct_clipped_display": "+100.00%",
1814
+ "improvement_pct_display": "+102.33%",
1815
+ "last": 2.6622845763340592,
1816
+ "late_mean": 2.49387572822161,
1817
+ "mid_mean": 2.022283688886091,
1818
+ "pressure": "pressure_up",
1819
+ "pressure_arrow": "\u2197",
1820
+ "pressure_display": "goal \u2191 | \u2197 +102.33% improving",
1821
+ "pressure_score": 1.0,
1822
+ "trend_label": "improving"
1823
+ },
1824
+ "val_structure_retrieval_loss": {
1825
+ "desired_direction": "lower",
1826
+ "early_mean": 11.562767315655947,
1827
+ "goal_display": "goal \u2193",
1828
+ "improvement_pct": 4.864921250164783,
1829
+ "improvement_pct_display": "+4.86%",
1830
+ "last": 10.882395941764116,
1831
+ "late_mean": 11.000247791409492,
1832
+ "mid_mean": 12.150514367967844,
1833
+ "pressure": "pressure_none",
1834
+ "pressure_arrow": "\u2192",
1835
+ "pressure_display": "goal \u2193 | \u2192 +4.86% flat",
1836
+ "trend_label": "flat"
1837
+ },
1838
+ "val_structure_target_rank_mean": {
1839
+ "desired_direction": "lower",
1840
+ "early_mean": 760.7540435791016,
1841
+ "goal_display": "goal \u2193",
1842
+ "improvement_pct": 45.44775026271755,
1843
+ "improvement_pct_display": "+45.45%",
1844
+ "last": 379.8781280517578,
1845
+ "late_mean": 415.0084457397461,
1846
+ "mid_mean": 683.3385162353516,
1847
+ "pressure": "pressure_up",
1848
+ "pressure_arrow": "\u2197",
1849
+ "pressure_display": "goal \u2193 | \u2197 +45.45% improving",
1850
+ "pressure_score": 1.0,
1851
+ "trend_label": "improving"
1852
+ },
1853
+ "val_structure_top1_top2_gap_mean": {
1854
+ "desired_direction": "higher",
1855
+ "early_mean": 0.07839679718017578,
1856
+ "goal_display": "goal \u2191",
1857
+ "improvement_pct": 54.90785232041847,
1858
+ "improvement_pct_display": "+54.91%",
1859
+ "last": 0.11855697631835938,
1860
+ "late_mean": 0.12144279479980469,
1861
+ "mid_mean": 0.1310901641845703,
1862
+ "pressure": "pressure_up",
1863
+ "pressure_arrow": "\u2197",
1864
+ "pressure_display": "goal \u2191 | \u2197 +54.91% improving",
1865
+ "pressure_score": 1.0,
1866
+ "trend_label": "improving"
1867
+ },
1868
+ "val_structure_top1_top2_gap_norm_mean": {
1869
+ "desired_direction": "higher",
1870
+ "early_mean": 0.06950851557485294,
1871
+ "goal_display": "goal \u2191",
1872
+ "improvement_pct": -16.826381575968767,
1873
+ "improvement_pct_display": "-16.83%",
1874
+ "last": 0.05411380449368153,
1875
+ "late_mean": 0.0578127475164365,
1876
+ "mid_mean": 0.07577372609375743,
1877
+ "pressure": "pressure_down",
1878
+ "pressure_arrow": "\u2198",
1879
+ "pressure_display": "goal \u2191 | \u2198 -16.83% worsening",
1880
+ "pressure_score": -1.0,
1881
+ "trend_label": "worsening"
1882
+ },
1883
+ "val_structure_top_negative_logit_mean": {
1884
+ "desired_direction": "lower",
1885
+ "early_mean": 9.2255859375,
1886
+ "goal_display": "goal \u2193",
1887
+ "improvement_pct": -16.07653223245475,
1888
+ "improvement_pct_display": "-16.08%",
1889
+ "last": 10.714599609375,
1890
+ "late_mean": 10.708740234375,
1891
+ "mid_mean": 10.51171875,
1892
+ "pressure": "pressure_down",
1893
+ "pressure_arrow": "\u2198",
1894
+ "pressure_display": "goal \u2193 | \u2198 -16.08% worsening",
1895
+ "pressure_score": -1.0,
1896
+ "trend_label": "worsening"
1897
+ },
1898
+ "val_structure_uncertain_fraction": {
1899
+ "desired_direction": "lower",
1900
+ "early_mean": 0.9941177368164062,
1901
+ "goal_display": "goal \u2193",
1902
+ "improvement_pct": -0.03261678728482514,
1903
+ "improvement_pct_display": "-0.03%",
1904
+ "last": 0.9922027587890625,
1905
+ "late_mean": 0.9944419860839844,
1906
+ "mid_mean": 0.9762382507324219,
1907
+ "pressure": "pressure_none",
1908
+ "pressure_arrow": "\u2192",
1909
+ "pressure_display": "goal \u2193 | \u2192 -0.03% flat",
1910
+ "trend_label": "flat"
1911
+ },
1912
+ "weighted_regularizer_to_contrastive_ratio": {
1913
+ "desired_direction": "lower",
1914
+ "early_mean": 0.0667222202939108,
1915
+ "goal_display": "goal \u2193",
1916
+ "improvement_pct": 54.60656419810926,
1917
+ "improvement_pct_display": "+54.61%",
1918
+ "last": 0.028996589606316913,
1919
+ "late_mean": 0.03028750823471251,
1920
+ "mid_mean": 0.0329437448262181,
1921
+ "pressure": "pressure_up",
1922
+ "pressure_arrow": "\u2197",
1923
+ "pressure_display": "goal \u2193 | \u2197 +54.61% improving",
1924
+ "pressure_score": 1.0,
1925
+ "trend_label": "improving"
1926
+ }
1927
+ },
1928
+ "metrics": {
1929
+ "backbone_lr": 3.886130904636371e-06,
1930
+ "checkpoint_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt",
1931
+ "checkpoint_reloaded": 1.0,
1932
+ "contrastive_loss": 0.5768923759460449,
1933
+ "covariance_loss": 10.012618064880371,
1934
+ "covariance_to_contrastive_ratio": 17.356128252623712,
1935
+ "covariance_weight": 0.0015,
1936
+ "dataloader_time_sec": 0.00035174899676349014,
1937
+ "dataloader_wait_pct_300s": 0.0382592069912683,
1938
+ "dataloader_wait_pct_60s": 0.0961140629935623,
1939
+ "dataloader_wait_pct_900s": 0.066640439237091,
1940
+ "dataloader_wait_pct_run": 0.06717802801040498,
1941
+ "different_inchikey_mean_cosine": 0.004679911238535274,
1942
+ "distributed_rank": 0.0,
1943
+ "distributed_world_size": 1.0,
1944
+ "effective_batch_size": 512.0,
1945
+ "effective_hard_negative_scale": 1.0,
1946
+ "effective_retrieval_weight": 0.6,
1947
+ "effective_structure_bce_weight": 1.0,
1948
+ "embedding_mean_norm": 1.0,
1949
+ "embedding_offdiag_mean_cosine": 0.009513536468148232,
1950
+ "embedding_std": 0.07216806709766388,
1951
+ "global_batch_size": 512.0,
1952
+ "gpu_active_fraction": 1.0,
1953
+ "gpu_count_observed": 2.0,
1954
+ "gpu_max_memory_gb": 13.279157161712646,
1955
+ "gpu_mem_used_mean_gb": 14.1943359375,
1956
+ "gpu_mem_util_mean_pct": 14.848754175733243,
1957
+ "gpu_memory_stable": true,
1958
+ "gpu_memory_utilization": 0.13982381154325935,
1959
+ "gpu_snapshot_available": 1.0,
1960
+ "gpu_util_max_pct": 54.0,
1961
+ "gpu_util_mean_pct": 36.0,
1962
+ "gpu_util_min_pct": 18.0,
1963
+ "grad_norm": 1.7659728527069092,
1964
+ "holdout_full_coverage": 0.0,
1965
+ "local_batch_size": 512.0,
1966
+ "lr": 3.886130904636371e-06,
1967
+ "mfu_estimate": 0.06267708747552383,
1968
+ "msms_objective": "hybrid",
1969
+ "optimizer_step": 1.0,
1970
+ "regularizer_ramp": 1.0,
1971
+ "resume_checkpoint_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt",
1972
+ "resume_start_step": 81002.0,
1973
+ "resumed_from_checkpoint": 1.0,
1974
+ "retrieval_head_lr": 1.9430654523181854e-05,
1975
+ "same_inchikey_mean_cosine": 0.16267980930137343,
1976
+ "samples_per_sec": 1399.0203442084614,
1977
+ "samples_per_sec_300s": 1547.033197547462,
1978
+ "samples_per_sec_60s": 1399.0203442084614,
1979
+ "samples_per_sec_900s": 1512.7885250414204,
1980
+ "samples_per_sec_run_mean": 1533.0231795943853,
1981
+ "step_time_cv_300s": 0.06743440395418321,
1982
+ "step_time_cv_60s": 0.0,
1983
+ "step_time_cv_900s": 0.09218780075615579,
1984
+ "step_time_sec": 0.3659703749981418,
1985
+ "structure_bce_locked": 0.0,
1986
+ "structure_bce_loss": 0.08383000642061234,
1987
+ "structure_candidate_count": 2126.0,
1988
+ "structure_clear_fraction": 0.0,
1989
+ "structure_clear_hit_at_1": 0.0,
1990
+ "structure_close_fraction": 0.029296875,
1991
+ "structure_close_hit_at_1": 0.0,
1992
+ "structure_confidence_bucket_score_mean": 0.10155132412910461,
1993
+ "structure_confidence_bucket_score_mode": "normalized_gap",
1994
+ "structure_confidence_candidate_count": 2126.0,
1995
+ "structure_confidence_clear_threshold": 1.0,
1996
+ "structure_confidence_close_threshold": 0.35,
1997
+ "structure_confidence_eval_key_mode": "target_identity",
1998
+ "structure_confidence_separation": -0.0040241447277367115,
1999
+ "structure_correct_logit_mean": 9.1875,
2000
+ "structure_fingerprint_cosine": 0.46021854877471924,
2001
+ "structure_hard_negative_bonus_applied": 1.0,
2002
+ "structure_hard_negative_bonus_mean": 0.060302734375,
2003
+ "structure_hard_negative_scale": 1.0,
2004
+ "structure_head_lr": 1.9430654523181854e-05,
2005
+ "structure_hit_at_1": 0.00390625,
2006
+ "structure_hit_at_20": 0.1171875,
2007
+ "structure_hit_at_5": 0.029296875,
2008
+ "structure_local_ce_loss": 3.545562744140625,
2009
+ "structure_local_ce_loss_weight": 0.4,
2010
+ "structure_local_rerank_loss": 0.0,
2011
+ "structure_local_rerank_loss_weight": 0.0,
2012
+ "structure_local_rerank_score_scale": 1.0,
2013
+ "structure_logit_margin_mean": -1.0078125,
2014
+ "structure_margin_loss": 0.0,
2015
+ "structure_margin_loss_weight": 0.0,
2016
+ "structure_missing_formula_fraction": 0.0,
2017
+ "structure_missing_inchikey_fraction": 0.0,
2018
+ "structure_pairwise_logistic_loss": 0.0,
2019
+ "structure_pairwise_logistic_loss_weight": 0.0,
2020
+ "structure_positive_count_mean": 1.0,
2021
+ "structure_post_label_candidate_count": 2126.0,
2022
+ "structure_post_precursor_candidate_count": 2128.0,
2023
+ "structure_prefilter_candidate_count": 2128.0,
2024
+ "structure_retrieval_logit_std": 0.9839304089546204,
2025
+ "structure_retrieval_loss": 10.630162239074707,
2026
+ "structure_retrieval_scorer": "projected_dot",
2027
+ "structure_retrieval_target_mode": "sample",
2028
+ "structure_skipped_query_fraction": 0.0,
2029
+ "structure_soft_target_exact_mix": 0.4,
2030
+ "structure_soft_target_loss": 7.523618221282959,
2031
+ "structure_soft_target_loss_weight": 0.3,
2032
+ "structure_target_backend": "rdkit_morgan",
2033
+ "structure_target_equivalent_count_mean": 1.0,
2034
+ "structure_target_identity_collision_count": 0.0,
2035
+ "structure_target_rank_mean": 314.703125,
2036
+ "structure_to_ssl_ratio": 217.7124787233957,
2037
+ "structure_top1_top2_gap_mean": 0.0986328125,
2038
+ "structure_top1_top2_gap_norm_mean": 0.10155132412910461,
2039
+ "structure_top_negative_logit_mean": 10.1875,
2040
+ "structure_uncertain_fraction": 0.970703125,
2041
+ "structure_uncertain_hit_at_1": 0.0040241447277367115,
2042
+ "structure_valid_query_fraction": 1.0,
2043
+ "test_expected_shard_count": 2.0,
2044
+ "test_full_coverage": 0.0,
2045
+ "test_observed_shard_count": 0.0,
2046
+ "train_loss": 6.491608619689941,
2047
+ "train_loss_ema": 6.519052214060582,
2048
+ "train_retrieval_bank_size": 4096.0,
2049
+ "val_deep_eval_single_view": 0.0,
2050
+ "val_deep_eval_structure_only": 0.0,
2051
+ "val_effective_structure_bce_weight": 1.0,
2052
+ "val_eval_loader_mode": "stream",
2053
+ "val_eval_mode": "deep",
2054
+ "val_expected_shard_count": 2.0,
2055
+ "val_full_coverage": 1.0,
2056
+ "val_hit_at_1": 0.0282135009765625,
2057
+ "val_hit_at_20": 0.15270233154296875,
2058
+ "val_hit_at_5": 0.0726776123046875,
2059
+ "val_loss": 6.646750174145564,
2060
+ "val_observed_eval_batches": 256.0,
2061
+ "val_observed_shard_count": 2.0,
2062
+ "val_ssl_loss": 1.0397093277424574,
2063
+ "val_structure_bce_loss": 0.06532714269997086,
2064
+ "val_structure_candidate_count": 4119.69140625,
2065
+ "val_structure_clear_fraction": 0.0,
2066
+ "val_structure_clear_hit_at_1": 0.0,
2067
+ "val_structure_close_fraction": 0.0077972412109375,
2068
+ "val_structure_close_hit_at_1": 0.014835291571216658,
2069
+ "val_structure_confidence_candidate_count": 4096.0,
2070
+ "val_structure_confidence_separation": -0.027957430157584895,
2071
+ "val_structure_correct_logit_mean": 9.49462890625,
2072
+ "val_structure_fingerprint_cosine": 0.4255251376889646,
2073
+ "val_structure_logit_margin_mean": -1.221923828125,
2074
+ "val_structure_missing_formula_fraction": 0.0,
2075
+ "val_structure_missing_inchikey_fraction": 0.0,
2076
+ "val_structure_retrieval_logit_std": 2.6622845763340592,
2077
+ "val_structure_retrieval_loss": 10.882395941764116,
2078
+ "val_structure_target_equivalent_count_mean": 1.0,
2079
+ "val_structure_target_identity_collision_count": 23.69140625,
2080
+ "val_structure_target_rank_mean": 379.8781280517578,
2081
+ "val_structure_top1_top2_gap_mean": 0.11855697631835938,
2082
+ "val_structure_top1_top2_gap_norm_mean": 0.05411380449368153,
2083
+ "val_structure_top_negative_logit_mean": 10.714599609375,
2084
+ "val_structure_uncertain_fraction": 0.9922027587890625,
2085
+ "val_structure_uncertain_hit_at_1": 0.027957430157584895,
2086
+ "variance_loss": 0.00213623046875,
2087
+ "variance_weight": 0.8,
2088
+ "weighted_covariance_loss": 0.015018927097320557,
2089
+ "weighted_regularizer_to_contrastive_ratio": 0.028996589606316913,
2090
+ "weighted_ssl_loss": 0.02968101501464844,
2091
+ "weighted_structure_bce_loss": 0.08383000642061234,
2092
+ "weighted_structure_retrieval_loss": 6.378097343444824,
2093
+ "weighted_variance_loss": 0.001708984375
2094
+ },
2095
+ "primary_metric_pressure_summary": {
2096
+ "average_improvement_pct": -14.247321653690681,
2097
+ "metric_count": 12,
2098
+ "overall_arrow": "\u2197",
2099
+ "overall_pressure": "pressure_up",
2100
+ "pressure_down_count": 4,
2101
+ "pressure_none_count": 3,
2102
+ "pressure_up_count": 5,
2103
+ "pressure_up_fraction": 0.4166666666666667,
2104
+ "weighted_improvement_pct": -4.446275812252011,
2105
+ "weighted_pressure_score": 0.1702127659574468
2106
+ },
2107
+ "primary_metric_pressure_summary_display": {
2108
+ "average_improvement_pct": -14.247321653690681,
2109
+ "average_improvement_pct_display": "-14.25%",
2110
+ "metric_count": 12,
2111
+ "overall_arrow": "\u2197",
2112
+ "overall_display": "\u2197 improving (-4.45% weighted)",
2113
+ "overall_label": "improving",
2114
+ "overall_pressure": "pressure_up",
2115
+ "pressure_down_count": 4,
2116
+ "pressure_none_count": 3,
2117
+ "pressure_up_count": 5,
2118
+ "pressure_up_fraction": 0.4166666666666667,
2119
+ "pressure_up_fraction_display": "+41.67%",
2120
+ "weighted_improvement_pct": -4.446275812252011,
2121
+ "weighted_improvement_pct_display": "-4.45%",
2122
+ "weighted_pressure_score": 0.1702127659574468
2123
+ },
2124
+ "primary_metric_pressures": {
2125
+ "contrastive_loss": {
2126
+ "desired_direction": "lower",
2127
+ "early_mean": 1.409792113166324,
2128
+ "improvement_pct": -1.6400841264380104,
2129
+ "improvement_pct_clipped": -1.6400841264380104,
2130
+ "last": 0.5768923759460449,
2131
+ "late_mean": 1.43291388983014,
2132
+ "mid_mean": 1.3598130204084982,
2133
+ "pressure": "pressure_none",
2134
+ "pressure_arrow": "\u2192",
2135
+ "pressure_score": 0.0
2136
+ },
2137
+ "grad_norm": {
2138
+ "desired_direction": "stable",
2139
+ "early_mean": 4.638080189338309,
2140
+ "improvement_pct": -100.0,
2141
+ "improvement_pct_clipped": -100.0,
2142
+ "last": 1.7659728527069092,
2143
+ "late_mean": 6.072961742165445,
2144
+ "mid_mean": 4.979553528603791,
2145
+ "pressure": "pressure_down",
2146
+ "pressure_arrow": "\u2198",
2147
+ "pressure_score": -1.0
2148
+ },
2149
+ "samples_per_sec": {
2150
+ "desired_direction": "higher",
2151
+ "early_mean": 1575.5664566260843,
2152
+ "improvement_pct": -2.8825714693532967,
2153
+ "improvement_pct_clipped": -2.8825714693532967,
2154
+ "last": 1399.0203442084614,
2155
+ "late_mean": 1530.1496274666802,
2156
+ "mid_mean": 1531.3735887168505,
2157
+ "pressure": "pressure_down",
2158
+ "pressure_arrow": "\u2198",
2159
+ "pressure_score": -1.0
2160
+ },
2161
+ "structure_bce_loss": {
2162
+ "desired_direction": "lower",
2163
+ "early_mean": 0.12293951650489272,
2164
+ "improvement_pct": 39.749156913940595,
2165
+ "improvement_pct_clipped": 39.749156913940595,
2166
+ "last": 0.08383000642061234,
2167
+ "late_mean": 0.07407209518012302,
2168
+ "mid_mean": 0.07652380578004556,
2169
+ "pressure": "pressure_up",
2170
+ "pressure_arrow": "\u2197",
2171
+ "pressure_score": 1.0
2172
+ },
2173
+ "train_loss": {
2174
+ "desired_direction": "lower",
2175
+ "early_mean": 6.340343519442343,
2176
+ "improvement_pct": 10.566866163080938,
2177
+ "improvement_pct_clipped": 10.566866163080938,
2178
+ "last": 6.491608619689941,
2179
+ "late_mean": 5.670367905463295,
2180
+ "mid_mean": 5.795929590401622,
2181
+ "pressure": "pressure_up",
2182
+ "pressure_arrow": "\u2197",
2183
+ "pressure_score": 1.0
2184
+ },
2185
+ "val_hit_at_20": {
2186
+ "desired_direction": "higher",
2187
+ "early_mean": 0.08158111572265625,
2188
+ "improvement_pct": 76.67165435331526,
2189
+ "improvement_pct_clipped": 76.67165435331526,
2190
+ "last": 0.15270233154296875,
2191
+ "late_mean": 0.14413070678710938,
2192
+ "mid_mean": 0.094146728515625,
2193
+ "pressure": "pressure_up",
2194
+ "pressure_arrow": "\u2197",
2195
+ "pressure_score": 1.0
2196
+ },
2197
+ "val_loss": {
2198
+ "desired_direction": "lower",
2199
+ "early_mean": 1.2876316868363775,
2200
+ "improvement_pct": -55.82459156154955,
2201
+ "improvement_pct_clipped": -55.82459156154955,
2202
+ "last": 6.646750174145564,
2203
+ "late_mean": 2.006446816829876,
2204
+ "mid_mean": 2.20697861268016,
2205
+ "pressure": "pressure_down",
2206
+ "pressure_arrow": "\u2198",
2207
+ "pressure_score": -1.0
2208
+ },
2209
+ "val_structure_bce_loss": {
2210
+ "desired_direction": "lower",
2211
+ "early_mean": 0.0771060927848642,
2212
+ "improvement_pct": 12.199713989530546,
2213
+ "improvement_pct_clipped": 12.199713989530546,
2214
+ "last": 0.06532714269997086,
2215
+ "late_mean": 0.06769936999660872,
2216
+ "mid_mean": 0.07029345531468391,
2217
+ "pressure": "pressure_up",
2218
+ "pressure_arrow": "\u2197",
2219
+ "pressure_score": 1.0
2220
+ },
2221
+ "val_structure_confidence_separation": {
2222
+ "desired_direction": "higher",
2223
+ "early_mean": -0.00795668972114072,
2224
+ "improvement_pct": -204.85499246505236,
2225
+ "improvement_pct_clipped": -100.0,
2226
+ "last": -0.027957430157584895,
2227
+ "late_mean": -0.024256365849851136,
2228
+ "mid_mean": -0.01309866098972634,
2229
+ "pressure": "pressure_down",
2230
+ "pressure_arrow": "\u2198",
2231
+ "pressure_score": -1.0
2232
+ },
2233
+ "val_structure_fingerprint_cosine": {
2234
+ "desired_direction": "higher",
2235
+ "early_mean": 0.4147872725734487,
2236
+ "improvement_pct": 0.17175282510408285,
2237
+ "improvement_pct_clipped": 0.17175282510408285,
2238
+ "last": 0.4255251376889646,
2239
+ "late_mean": 0.4154996814322658,
2240
+ "mid_mean": 0.4204998221830465,
2241
+ "pressure": "pressure_none",
2242
+ "pressure_arrow": "\u2192",
2243
+ "pressure_score": 0.0
2244
+ },
2245
+ "val_structure_top1_top2_gap_mean": {
2246
+ "desired_direction": "higher",
2247
+ "early_mean": 0.07839679718017578,
2248
+ "improvement_pct": 54.90785232041847,
2249
+ "improvement_pct_clipped": 54.90785232041847,
2250
+ "last": 0.11855697631835938,
2251
+ "late_mean": 0.12144279479980469,
2252
+ "mid_mean": 0.1310901641845703,
2253
+ "pressure": "pressure_up",
2254
+ "pressure_arrow": "\u2197",
2255
+ "pressure_score": 1.0
2256
+ },
2257
+ "val_structure_uncertain_fraction": {
2258
+ "desired_direction": "lower",
2259
+ "early_mean": 0.9941177368164062,
2260
+ "improvement_pct": -0.03261678728482514,
2261
+ "improvement_pct_clipped": -0.03261678728482514,
2262
+ "last": 0.9922027587890625,
2263
+ "late_mean": 0.9944419860839844,
2264
+ "mid_mean": 0.9762382507324219,
2265
+ "pressure": "pressure_none",
2266
+ "pressure_arrow": "\u2192",
2267
+ "pressure_score": 0.0
2268
+ }
2269
+ },
2270
+ "primary_metric_pressures_display": {
2271
+ "contrastive_loss": {
2272
+ "desired_direction": "lower",
2273
+ "early_mean": 1.409792113166324,
2274
+ "goal_display": "goal \u2193",
2275
+ "improvement_pct": -1.6400841264380104,
2276
+ "improvement_pct_display": "-1.64%",
2277
+ "last": 0.5768923759460449,
2278
+ "late_mean": 1.43291388983014,
2279
+ "mid_mean": 1.3598130204084982,
2280
+ "pressure": "pressure_none",
2281
+ "pressure_arrow": "\u2192",
2282
+ "pressure_display": "goal \u2193 | \u2192 -1.64% flat",
2283
+ "trend_label": "flat"
2284
+ },
2285
+ "grad_norm": {
2286
+ "desired_direction": "stable",
2287
+ "early_mean": 4.638080189338309,
2288
+ "goal_display": "goal \u2248",
2289
+ "improvement_pct": -100.0,
2290
+ "improvement_pct_display": "-100.00%",
2291
+ "last": 1.7659728527069092,
2292
+ "late_mean": 6.072961742165445,
2293
+ "mid_mean": 4.979553528603791,
2294
+ "pressure": "pressure_down",
2295
+ "pressure_arrow": "\u2198",
2296
+ "pressure_display": "goal \u2248 | \u2198 -100.00% worsening",
2297
+ "pressure_score": -1.0,
2298
+ "trend_label": "worsening"
2299
+ },
2300
+ "samples_per_sec": {
2301
+ "desired_direction": "higher",
2302
+ "early_mean": 1575.5664566260843,
2303
+ "goal_display": "goal \u2191",
2304
+ "improvement_pct": -2.8825714693532967,
2305
+ "improvement_pct_display": "-2.88%",
2306
+ "last": 1399.0203442084614,
2307
+ "late_mean": 1530.1496274666802,
2308
+ "mid_mean": 1531.3735887168505,
2309
+ "pressure": "pressure_down",
2310
+ "pressure_arrow": "\u2198",
2311
+ "pressure_display": "goal \u2191 | \u2198 -2.88% worsening",
2312
+ "pressure_score": -1.0,
2313
+ "trend_label": "worsening"
2314
+ },
2315
+ "structure_bce_loss": {
2316
+ "desired_direction": "lower",
2317
+ "early_mean": 0.12293951650489272,
2318
+ "goal_display": "goal \u2193",
2319
+ "improvement_pct": 39.749156913940595,
2320
+ "improvement_pct_display": "+39.75%",
2321
+ "last": 0.08383000642061234,
2322
+ "late_mean": 0.07407209518012302,
2323
+ "mid_mean": 0.07652380578004556,
2324
+ "pressure": "pressure_up",
2325
+ "pressure_arrow": "\u2197",
2326
+ "pressure_display": "goal \u2193 | \u2197 +39.75% improving",
2327
+ "pressure_score": 1.0,
2328
+ "trend_label": "improving"
2329
+ },
2330
+ "train_loss": {
2331
+ "desired_direction": "lower",
2332
+ "early_mean": 6.340343519442343,
2333
+ "goal_display": "goal \u2193",
2334
+ "improvement_pct": 10.566866163080938,
2335
+ "improvement_pct_display": "+10.57%",
2336
+ "last": 6.491608619689941,
2337
+ "late_mean": 5.670367905463295,
2338
+ "mid_mean": 5.795929590401622,
2339
+ "pressure": "pressure_up",
2340
+ "pressure_arrow": "\u2197",
2341
+ "pressure_display": "goal \u2193 | \u2197 +10.57% improving",
2342
+ "pressure_score": 1.0,
2343
+ "trend_label": "improving"
2344
+ },
2345
+ "val_hit_at_20": {
2346
+ "desired_direction": "higher",
2347
+ "early_mean": 0.08158111572265625,
2348
+ "goal_display": "goal \u2191",
2349
+ "improvement_pct": 76.67165435331526,
2350
+ "improvement_pct_display": "+76.67%",
2351
+ "last": 0.15270233154296875,
2352
+ "late_mean": 0.14413070678710938,
2353
+ "mid_mean": 0.094146728515625,
2354
+ "pressure": "pressure_up",
2355
+ "pressure_arrow": "\u2197",
2356
+ "pressure_display": "goal \u2191 | \u2197 +76.67% improving",
2357
+ "pressure_score": 1.0,
2358
+ "trend_label": "improving"
2359
+ },
2360
+ "val_loss": {
2361
+ "desired_direction": "lower",
2362
+ "early_mean": 1.2876316868363775,
2363
+ "goal_display": "goal \u2193",
2364
+ "improvement_pct": -55.82459156154955,
2365
+ "improvement_pct_display": "-55.82%",
2366
+ "last": 6.646750174145564,
2367
+ "late_mean": 2.006446816829876,
2368
+ "mid_mean": 2.20697861268016,
2369
+ "pressure": "pressure_down",
2370
+ "pressure_arrow": "\u2198",
2371
+ "pressure_display": "goal \u2193 | \u2198 -55.82% worsening",
2372
+ "pressure_score": -1.0,
2373
+ "trend_label": "worsening"
2374
+ },
2375
+ "val_structure_bce_loss": {
2376
+ "desired_direction": "lower",
2377
+ "early_mean": 0.0771060927848642,
2378
+ "goal_display": "goal \u2193",
2379
+ "improvement_pct": 12.199713989530546,
2380
+ "improvement_pct_display": "+12.20%",
2381
+ "last": 0.06532714269997086,
2382
+ "late_mean": 0.06769936999660872,
2383
+ "mid_mean": 0.07029345531468391,
2384
+ "pressure": "pressure_up",
2385
+ "pressure_arrow": "\u2197",
2386
+ "pressure_display": "goal \u2193 | \u2197 +12.20% improving",
2387
+ "pressure_score": 1.0,
2388
+ "trend_label": "improving"
2389
+ },
2390
+ "val_structure_confidence_separation": {
2391
+ "desired_direction": "higher",
2392
+ "early_mean": -0.00795668972114072,
2393
+ "goal_display": "goal \u2191",
2394
+ "improvement_pct": -204.85499246505236,
2395
+ "improvement_pct_clipped": -100.0,
2396
+ "improvement_pct_clipped_display": "-100.00%",
2397
+ "improvement_pct_display": "-204.85%",
2398
+ "last": -0.027957430157584895,
2399
+ "late_mean": -0.024256365849851136,
2400
+ "mid_mean": -0.01309866098972634,
2401
+ "pressure": "pressure_down",
2402
+ "pressure_arrow": "\u2198",
2403
+ "pressure_display": "goal \u2191 | \u2198 -204.85% worsening",
2404
+ "pressure_score": -1.0,
2405
+ "trend_label": "worsening"
2406
+ },
2407
+ "val_structure_fingerprint_cosine": {
2408
+ "desired_direction": "higher",
2409
+ "early_mean": 0.4147872725734487,
2410
+ "goal_display": "goal \u2191",
2411
+ "improvement_pct": 0.17175282510408285,
2412
+ "improvement_pct_display": "+0.17%",
2413
+ "last": 0.4255251376889646,
2414
+ "late_mean": 0.4154996814322658,
2415
+ "mid_mean": 0.4204998221830465,
2416
+ "pressure": "pressure_none",
2417
+ "pressure_arrow": "\u2192",
2418
+ "pressure_display": "goal \u2191 | \u2192 +0.17% flat",
2419
+ "trend_label": "flat"
2420
+ },
2421
+ "val_structure_top1_top2_gap_mean": {
2422
+ "desired_direction": "higher",
2423
+ "early_mean": 0.07839679718017578,
2424
+ "goal_display": "goal \u2191",
2425
+ "improvement_pct": 54.90785232041847,
2426
+ "improvement_pct_display": "+54.91%",
2427
+ "last": 0.11855697631835938,
2428
+ "late_mean": 0.12144279479980469,
2429
+ "mid_mean": 0.1310901641845703,
2430
+ "pressure": "pressure_up",
2431
+ "pressure_arrow": "\u2197",
2432
+ "pressure_display": "goal \u2191 | \u2197 +54.91% improving",
2433
+ "pressure_score": 1.0,
2434
+ "trend_label": "improving"
2435
+ },
2436
+ "val_structure_uncertain_fraction": {
2437
+ "desired_direction": "lower",
2438
+ "early_mean": 0.9941177368164062,
2439
+ "goal_display": "goal \u2193",
2440
+ "improvement_pct": -0.03261678728482514,
2441
+ "improvement_pct_display": "-0.03%",
2442
+ "last": 0.9922027587890625,
2443
+ "late_mean": 0.9944419860839844,
2444
+ "mid_mean": 0.9762382507324219,
2445
+ "pressure": "pressure_none",
2446
+ "pressure_arrow": "\u2192",
2447
+ "pressure_display": "goal \u2193 | \u2192 -0.03% flat",
2448
+ "trend_label": "flat"
2449
+ }
2450
+ },
2451
+ "primary_metrics": {
2452
+ "contrastive_loss": 0.5768923759460449,
2453
+ "different_inchikey_mean_cosine": 0.004679911238535274,
2454
+ "embedding_std": 0.07216806709766388,
2455
+ "grad_norm": 1.7659728527069092,
2456
+ "samples_per_sec": 1399.0203442084614,
2457
+ "structure_bce_loss": 0.08383000642061234,
2458
+ "train_loss": 6.491608619689941,
2459
+ "val_hit_at_20": 0.15270233154296875,
2460
+ "val_loss": 6.646750174145564,
2461
+ "val_structure_bce_loss": 0.06532714269997086,
2462
+ "val_structure_candidate_count": 4119.69140625,
2463
+ "val_structure_clear_hit_at_1": 0.0,
2464
+ "val_structure_confidence_separation": -0.027957430157584895,
2465
+ "val_structure_fingerprint_cosine": 0.4255251376889646,
2466
+ "val_structure_top1_top2_gap_mean": 0.11855697631835938,
2467
+ "val_structure_uncertain_fraction": 0.9922027587890625
2468
+ },
2469
+ "run_name": "final_v3_structure_alignment_v26_r1",
2470
+ "shard_counts": {
2471
+ "test": 2,
2472
+ "train": 4,
2473
+ "val": 2
2474
+ }
2475
+ }
metadata/archive/20260413T213902Z/latest_checkpoint.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ /home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt
metadata/archive/20260413T213902Z/latest_metrics.json ADDED
@@ -0,0 +1,169 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backbone_lr": 3.886130904636371e-06,
3
+ "checkpoint_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt",
4
+ "checkpoint_reloaded": 1.0,
5
+ "contrastive_loss": 0.5768923759460449,
6
+ "covariance_loss": 10.012618064880371,
7
+ "covariance_to_contrastive_ratio": 17.356128252623712,
8
+ "covariance_weight": 0.0015,
9
+ "dataloader_time_sec": 0.00035174899676349014,
10
+ "dataloader_wait_pct_300s": 0.0382592069912683,
11
+ "dataloader_wait_pct_60s": 0.0961140629935623,
12
+ "dataloader_wait_pct_900s": 0.066640439237091,
13
+ "dataloader_wait_pct_run": 0.06717802801040498,
14
+ "different_inchikey_mean_cosine": 0.004679911238535274,
15
+ "distributed_rank": 0.0,
16
+ "distributed_world_size": 1.0,
17
+ "effective_batch_size": 512.0,
18
+ "effective_hard_negative_scale": 1.0,
19
+ "effective_retrieval_weight": 0.6,
20
+ "effective_structure_bce_weight": 1.0,
21
+ "embedding_mean_norm": 1.0,
22
+ "embedding_offdiag_mean_cosine": 0.009513536468148232,
23
+ "embedding_std": 0.07216806709766388,
24
+ "final_report": 1.0,
25
+ "global_batch_size": 512.0,
26
+ "gpu_active_fraction": 1.0,
27
+ "gpu_count_observed": 2.0,
28
+ "gpu_max_memory_gb": 13.279157161712646,
29
+ "gpu_mem_used_mean_gb": 14.1943359375,
30
+ "gpu_mem_util_mean_pct": 14.848754175733243,
31
+ "gpu_memory_stable": true,
32
+ "gpu_memory_utilization": 0.13982381154325935,
33
+ "gpu_snapshot_available": 1.0,
34
+ "gpu_util_max_pct": 54.0,
35
+ "gpu_util_mean_pct": 36.0,
36
+ "gpu_util_min_pct": 18.0,
37
+ "grad_norm": 1.7659728527069092,
38
+ "holdout_full_coverage": 0.0,
39
+ "local_batch_size": 512.0,
40
+ "lr": 3.886130904636371e-06,
41
+ "mfu_estimate": 0.06267708747552383,
42
+ "msms_objective": "hybrid",
43
+ "optimizer_step": 1.0,
44
+ "regularizer_ramp": 1.0,
45
+ "resume_checkpoint_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt",
46
+ "resume_start_step": 81002.0,
47
+ "resumed_from_checkpoint": 1.0,
48
+ "retrieval_head_lr": 1.9430654523181854e-05,
49
+ "same_inchikey_mean_cosine": 0.16267980930137343,
50
+ "samples_per_sec": 1399.0203442084614,
51
+ "samples_per_sec_300s": 1547.033197547462,
52
+ "samples_per_sec_60s": 1399.0203442084614,
53
+ "samples_per_sec_900s": 1512.7885250414204,
54
+ "samples_per_sec_run_mean": 1533.0231795943853,
55
+ "step": 91001,
56
+ "step_time_cv_300s": 0.06743440395418321,
57
+ "step_time_cv_60s": 0.0,
58
+ "step_time_cv_900s": 0.09218780075615579,
59
+ "step_time_sec": 0.3659703749981418,
60
+ "structure_bce_locked": 0.0,
61
+ "structure_bce_loss": 0.08383000642061234,
62
+ "structure_candidate_count": 2126.0,
63
+ "structure_clear_fraction": 0.0,
64
+ "structure_clear_hit_at_1": 0.0,
65
+ "structure_close_fraction": 0.029296875,
66
+ "structure_close_hit_at_1": 0.0,
67
+ "structure_confidence_bucket_score_mean": 0.10155132412910461,
68
+ "structure_confidence_bucket_score_mode": "normalized_gap",
69
+ "structure_confidence_candidate_count": 2126.0,
70
+ "structure_confidence_clear_threshold": 1.0,
71
+ "structure_confidence_close_threshold": 0.35,
72
+ "structure_confidence_eval_key_mode": "target_identity",
73
+ "structure_confidence_separation": -0.0040241447277367115,
74
+ "structure_correct_logit_mean": 9.1875,
75
+ "structure_fingerprint_cosine": 0.46021854877471924,
76
+ "structure_hard_negative_bonus_applied": 1.0,
77
+ "structure_hard_negative_bonus_mean": 0.060302734375,
78
+ "structure_hard_negative_scale": 1.0,
79
+ "structure_head_lr": 1.9430654523181854e-05,
80
+ "structure_hit_at_1": 0.00390625,
81
+ "structure_hit_at_20": 0.1171875,
82
+ "structure_hit_at_5": 0.029296875,
83
+ "structure_local_ce_loss": 3.545562744140625,
84
+ "structure_local_ce_loss_weight": 0.4,
85
+ "structure_local_rerank_loss": 0.0,
86
+ "structure_local_rerank_loss_weight": 0.0,
87
+ "structure_local_rerank_score_scale": 1.0,
88
+ "structure_logit_margin_mean": -1.0078125,
89
+ "structure_margin_loss": 0.0,
90
+ "structure_margin_loss_weight": 0.0,
91
+ "structure_missing_formula_fraction": 0.0,
92
+ "structure_missing_inchikey_fraction": 0.0,
93
+ "structure_pairwise_logistic_loss": 0.0,
94
+ "structure_pairwise_logistic_loss_weight": 0.0,
95
+ "structure_positive_count_mean": 1.0,
96
+ "structure_post_label_candidate_count": 2126.0,
97
+ "structure_post_precursor_candidate_count": 2128.0,
98
+ "structure_prefilter_candidate_count": 2128.0,
99
+ "structure_retrieval_logit_std": 0.9839304089546204,
100
+ "structure_retrieval_loss": 10.630162239074707,
101
+ "structure_retrieval_scorer": "projected_dot",
102
+ "structure_retrieval_target_mode": "sample",
103
+ "structure_skipped_query_fraction": 0.0,
104
+ "structure_soft_target_exact_mix": 0.4,
105
+ "structure_soft_target_loss": 7.523618221282959,
106
+ "structure_soft_target_loss_weight": 0.3,
107
+ "structure_target_backend": "rdkit_morgan",
108
+ "structure_target_equivalent_count_mean": 1.0,
109
+ "structure_target_identity_collision_count": 0.0,
110
+ "structure_target_rank_mean": 314.703125,
111
+ "structure_to_ssl_ratio": 217.7124787233957,
112
+ "structure_top1_top2_gap_mean": 0.0986328125,
113
+ "structure_top1_top2_gap_norm_mean": 0.10155132412910461,
114
+ "structure_top_negative_logit_mean": 10.1875,
115
+ "structure_uncertain_fraction": 0.970703125,
116
+ "structure_uncertain_hit_at_1": 0.0040241447277367115,
117
+ "structure_valid_query_fraction": 1.0,
118
+ "test_expected_shard_count": 2.0,
119
+ "test_full_coverage": 0.0,
120
+ "test_observed_shard_count": 0.0,
121
+ "train_loss": 6.491608619689941,
122
+ "train_loss_ema": 6.519052214060582,
123
+ "train_retrieval_bank_size": 4096.0,
124
+ "val_deep_eval_single_view": 0.0,
125
+ "val_deep_eval_structure_only": 0.0,
126
+ "val_effective_structure_bce_weight": 1.0,
127
+ "val_eval_loader_mode": "stream",
128
+ "val_eval_mode": "deep",
129
+ "val_expected_shard_count": 2.0,
130
+ "val_full_coverage": 1.0,
131
+ "val_hit_at_1": 0.0282135009765625,
132
+ "val_hit_at_20": 0.15270233154296875,
133
+ "val_hit_at_5": 0.0726776123046875,
134
+ "val_loss": 6.646750174145564,
135
+ "val_observed_eval_batches": 256.0,
136
+ "val_observed_shard_count": 2.0,
137
+ "val_ssl_loss": 1.0397093277424574,
138
+ "val_structure_bce_loss": 0.06532714269997086,
139
+ "val_structure_candidate_count": 4119.69140625,
140
+ "val_structure_clear_fraction": 0.0,
141
+ "val_structure_clear_hit_at_1": 0.0,
142
+ "val_structure_close_fraction": 0.0077972412109375,
143
+ "val_structure_close_hit_at_1": 0.014835291571216658,
144
+ "val_structure_confidence_candidate_count": 4096.0,
145
+ "val_structure_confidence_separation": -0.027957430157584895,
146
+ "val_structure_correct_logit_mean": 9.49462890625,
147
+ "val_structure_fingerprint_cosine": 0.4255251376889646,
148
+ "val_structure_logit_margin_mean": -1.221923828125,
149
+ "val_structure_missing_formula_fraction": 0.0,
150
+ "val_structure_missing_inchikey_fraction": 0.0,
151
+ "val_structure_retrieval_logit_std": 2.6622845763340592,
152
+ "val_structure_retrieval_loss": 10.882395941764116,
153
+ "val_structure_target_equivalent_count_mean": 1.0,
154
+ "val_structure_target_identity_collision_count": 23.69140625,
155
+ "val_structure_target_rank_mean": 379.8781280517578,
156
+ "val_structure_top1_top2_gap_mean": 0.11855697631835938,
157
+ "val_structure_top1_top2_gap_norm_mean": 0.05411380449368153,
158
+ "val_structure_top_negative_logit_mean": 10.714599609375,
159
+ "val_structure_uncertain_fraction": 0.9922027587890625,
160
+ "val_structure_uncertain_hit_at_1": 0.027957430157584895,
161
+ "variance_loss": 0.00213623046875,
162
+ "variance_weight": 0.8,
163
+ "weighted_covariance_loss": 0.015018927097320557,
164
+ "weighted_regularizer_to_contrastive_ratio": 0.028996589606316913,
165
+ "weighted_ssl_loss": 0.02968101501464844,
166
+ "weighted_structure_bce_loss": 0.08383000642061234,
167
+ "weighted_structure_retrieval_loss": 6.378097343444824,
168
+ "weighted_variance_loss": 0.001708984375
169
+ }
metadata/archive/20260413T213902Z/perf_summary.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "logged_points": 521,
3
+ "train_loss_min": 3.2278988361358643,
4
+ "train_loss_max": 7.677141189575195,
5
+ "train_loss_last": 6.491608619689941,
6
+ "train_loss_ema_last": 6.519052214060582,
7
+ "weighted_regularizer_ratio_median": 0.02686493425176378,
8
+ "weighted_regularizer_ratio_p90": 0.09558363904110317,
9
+ "samples_per_sec_mean": 1545.3851888056406,
10
+ "samples_per_sec_min": 930.5498563655322,
11
+ "samples_per_sec_max": 2247.460787425693,
12
+ "step_time_mean": 0.33468720350667036,
13
+ "mfu_mean": 0.06923433462788683,
14
+ "mfu_max": 0.10068781126337953,
15
+ "gpu_mem_max_gb": 13.279157161712646
16
+ }
metadata/archive/20260413T213902Z/results.md ADDED
@@ -0,0 +1,171 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # MS/MS Training Run Result
2
+
3
+ - Run name: `final_v3_structure_alignment_v26_r1`
4
+ - Created at: `2026-04-13T21:36:19Z`
5
+ - Dataset root: `/home/ubuntu/data/gems_v1_handoff_labeled_v26`
6
+ - Shard counts: `{'train': 4, 'val': 2, 'test': 2}`
7
+
8
+ ## Primary Metrics
9
+
10
+ - `train_loss`: `6.491608619689941`
11
+ - `val_loss`: `6.646750174145564`
12
+ - `grad_norm`: `1.7659728527069092`
13
+ - `contrastive_loss`: `0.5768923759460449`
14
+ - `structure_bce_loss`: `0.08383000642061234`
15
+ - `val_structure_bce_loss`: `0.06532714269997086`
16
+ - `val_structure_fingerprint_cosine`: `0.4255251376889646`
17
+ - `different_inchikey_mean_cosine`: `0.004679911238535274`
18
+ - `embedding_std`: `0.07216806709766388`
19
+ - `val_hit_at_20`: `0.15270233154296875`
20
+ - `val_structure_candidate_count`: `4119.69140625`
21
+ - `val_structure_top1_top2_gap_mean`: `0.11855697631835938`
22
+ - `val_structure_uncertain_fraction`: `0.9922027587890625`
23
+ - `val_structure_clear_hit_at_1`: `0.0`
24
+ - `val_structure_confidence_separation`: `-0.027957430157584895`
25
+ - `samples_per_sec`: `1399.0203442084614`
26
+
27
+ ## Failure Mode Assessment
28
+
29
+ - `dataset_shards_accessible`: pass. Observed shard counts {'train': 4, 'val': 2, 'test': 2}; the canonical Day 1 loader contract requires all splits to be readable.
30
+ - `dataloader_streaming_stable`: pass. Training batches continued to arrive with positive throughput and no iterator stall.
31
+ - `non_finite_loss_or_tensor_failure`: pass. Technical_Plan requires zero NaN or Inf tensors and finite train/val loss.
32
+ - `validation_observability_gap`: pass. Engineering.md expects slope-aware monitoring; missing validation or retrieval metrics make trend judgment weaker.
33
+ - `gpu_memory_instability`: pass. Technical_Plan Day 1 success requires stable GPU memory through the bounded run.
34
+ - `checkpoint_save_reload_failure`: pass. Checkpoint save and reload is an explicit Day 1 success criterion.
35
+ - `embedding_collapse`: pass. Engineering.md expects embedding variance monitoring and non-collapsed representations.
36
+ - `regularizer_dominance`: pass. The current instability mode is regularization overpowering the SSL objective and producing jagged total loss.
37
+ - `structure_supervision_gap`: pass. Once the bridge phase begins, Technical_Plan requires retrieval-oriented validation rather than only augmented contrastive loss.
38
+ - `confidence_surface_gap`: pass. Structure-first qualification now requires explicit confidence diagnostics such as top-1/top-2 separation and uncertainty buckets, not just broader retrieval metrics.
39
+ - `chemistry_target_backend_gap`: pass. The bridge phase should prefer chemistry-grounded fingerprint targets; hashed fallback targets weaken the retrieval signal.
40
+ - `bridge_fit_without_ranking`: pass. Bridge runs must improve retrieval after fingerprint fit saturates; fitting BCE while val Hit@20 stays at candidate-bank baseline means ranking is not generalizing.
41
+ - `metadata_shortcutting`: watch. Engineering.md flags acquisition-artefact shortcut risk; adduct and instrument embeddings remain deliberately down-scaled.
42
+
43
+ ## Pressure Summary
44
+
45
+ - all metrics overall: `↗ improving (+21.89% weighted)`
46
+ - all metrics weighted improvement: `+21.89%`
47
+ - all metrics average improvement: `+34.52%`
48
+ - all metrics weighted pressure score: `0.284`
49
+ - all metrics counts: `up=21` `none=7` `down=11` `up_fraction=+53.85%`
50
+ - primary metrics overall: `↗ improving (-4.45% weighted)`
51
+ - primary metrics weighted improvement: `-4.45%`
52
+
53
+ ## Metric Pressure
54
+
55
+ - `train_loss`: `goal ↓ | ↗ +10.57% improving` (early `6.340343519442343`, mid `5.795929590401622`, late `5.670367905463295`, last `6.491608619689941`)
56
+ - `train_loss_ema`: `goal ↓ | ↗ +10.46% improving` (early `6.265581717362126`, mid `5.8371378655061745`, late `5.610470838300897`, last `6.519052214060582`)
57
+ - `contrastive_loss`: `goal ↓ | → -1.64% flat` (early `1.409792113166324`, mid `1.3598130204084982`, late `1.43291388983014`, last `0.5768923759460449`)
58
+ - `val_loss`: `goal ↓ | ↘ -55.82% worsening` (early `1.2876316868363775`, mid `2.20697861268016`, late `2.006446816829876`, last `6.646750174145564`)
59
+ - `val_ssl_loss`: `goal ↓ | → +2.32% flat` (early `1.0849772497183745`, mid `1.0747967428483425`, late `1.0598136518327963`, last `1.0397093277424574`)
60
+ - `structure_bce_loss`: `goal ↓ | ↗ +39.75% improving` (early `0.12293951650489272`, mid `0.07652380578004556`, late `0.07407209518012302`, last `0.08383000642061234`)
61
+ - `val_structure_bce_loss`: `goal ↓ | ↗ +12.20% improving` (early `0.0771060927848642`, mid `0.07029345531468391`, late `0.06769936999660872`, last `0.06532714269997086`)
62
+ - `structure_retrieval_loss`: `goal ↓ | ↗ +10.09% improving` (early `10.237479978903181`, mid `9.41523380775672`, late `9.204101929719421`, last `10.630162239074707`)
63
+ - `val_structure_retrieval_loss`: `goal ↓ | → +4.86% flat` (early `11.562767315655947`, mid `12.150514367967844`, late `11.000247791409492`, last `10.882395941764116`)
64
+ - `structure_fingerprint_cosine`: `goal ↑ | ↗ +17.59% improving` (early `0.3965519580968543`, mid `0.4464927057310336`, late `0.46629545006943846`, last `0.46021854877471924`)
65
+ - `val_structure_fingerprint_cosine`: `goal ↑ | → +0.17% flat` (early `0.4147872725734487`, mid `0.4204998221830465`, late `0.4154996814322658`, last `0.4255251376889646`)
66
+ - `structure_target_rank_mean`: `goal ↓ | ↗ +58.80% improving` (early `351.84337743858384`, mid `176.16389315390174`, late `144.9534280711207`, last `314.703125`)
67
+ - `val_structure_target_rank_mean`: `goal ↓ | ↗ +45.45% improving` (early `760.7540435791016`, mid `683.3385162353516`, late `415.0084457397461`, last `379.8781280517578`)
68
+ - `structure_correct_logit_mean`: `goal ↑ | ↗ +47.31% improving` (early `6.134031791907514`, mid `9.474891618497109`, late `9.035739942528735`, last `9.1875`)
69
+ - `val_structure_correct_logit_mean`: `goal ↑ | ↗ +16.44% improving` (early `8.0311279296875`, mid `8.441162109375`, late `9.3515625`, last `9.49462890625`)
70
+ - `structure_top_negative_logit_mean`: `goal ↓ | ↘ -39.72% worsening` (early `7.262102601156069`, mid `10.490606936416185`, late `10.14691091954023`, last `10.1875`)
71
+ - `val_structure_top_negative_logit_mean`: `goal ↓ | ↘ -16.08% worsening` (early `9.2255859375`, mid `10.51171875`, late `10.708740234375`, last `10.714599609375`)
72
+ - `structure_logit_margin_mean`: `goal ↑ | → +1.23% flat` (early `-1.1263999277456647`, mid `-1.0154725885115607`, late `-1.1125740840517242`, last `-1.0078125`)
73
+ - `val_structure_logit_margin_mean`: `goal ↑ | ↘ -13.67% worsening` (early `-1.194091796875`, mid `-2.0707931518554688`, late `-1.3573150634765625`, last `-1.221923828125`)
74
+ - `structure_retrieval_logit_std`: `goal ↑ | ↗ +112.06% improving` (early `1.0455416592559375`, mid `2.0036371215919537`, late `2.2171744848119803`, last `0.9839304089546204`)
75
+ - `val_structure_retrieval_logit_std`: `goal ↑ | ↗ +102.33% improving` (early `1.2326001720502973`, mid `2.022283688886091`, late `2.49387572822161`, last `2.6622845763340592`)
76
+ - `structure_top1_top2_gap_mean`: `goal ↑ | ↗ +59.22% improving` (early `0.09793567106213873`, mid `0.12284365968208093`, late `0.1559343009159483`, last `0.0986328125`)
77
+ - `val_structure_top1_top2_gap_mean`: `goal ↑ | ↗ +54.91% improving` (early `0.07839679718017578`, mid `0.1310901641845703`, late `0.12144279479980469`, last `0.11855697631835938`)
78
+ - `structure_top1_top2_gap_norm_mean`: `goal ↑ | ↘ -35.65% worsening` (early `0.12215662799301864`, mid `0.07032911504236604`, late `0.07860379065547524`, last `0.10155132412910461`)
79
+ - `val_structure_top1_top2_gap_norm_mean`: `goal ↑ | ↘ -16.83% worsening` (early `0.06950851557485294`, mid `0.07577372609375743`, late `0.0578127475164365`, last `0.05411380449368153`)
80
+ - `structure_uncertain_fraction`: `goal ↓ | ↘ -6.11% worsening` (early `0.924663565751445`, mid `0.984747561416185`, late `0.9811422413793104`, last `0.970703125`)
81
+ - `val_structure_uncertain_fraction`: `goal ↓ | → -0.03% flat` (early `0.9941177368164062`, mid `0.9762382507324219`, late `0.9944419860839844`, last `0.9922027587890625`)
82
+ - `structure_confidence_separation`: `goal ↑ | ↘ -531.56% worsening` (early `-0.007918493679557749`, mid `-0.032136323077151974`, late `-0.05001017991881603`, last `-0.0040241447277367115`)
83
+ - `val_structure_confidence_separation`: `goal ↑ | ↘ -204.85% worsening` (early `-0.00795668972114072`, mid `-0.01309866098972634`, late `-0.024256365849851136`, last `-0.027957430157584895`)
84
+ - `structure_hit_at_1`: `goal ↑ | ↗ +549.68% improving` (early `0.007880238439306358`, mid `0.032706376445086706`, late `0.051196569683908046`, last `0.00390625`)
85
+ - `structure_hit_at_5`: `goal ↑ | ↗ +402.47% improving` (early `0.03360955563583815`, mid `0.11673591040462428`, late `0.16887796336206898`, last `0.029296875`)
86
+ - `structure_hit_at_20`: `goal ↑ | ↗ +239.46% improving` (early `0.10721865968208093`, mid `0.28105807442196534`, late `0.3639659662356322`, last `0.1171875`)
87
+ - `val_hit_at_1`: `goal ↑ | → +207.79% flat` (early `0.0079345703125`, mid `0.01372528076171875`, late `0.02442169189453125`, last `0.0282135009765625`)
88
+ - `val_hit_at_5`: `goal ↑ | ↗ +134.85% improving` (early `0.0288543701171875`, mid `0.0425567626953125`, late `0.0677642822265625`, last `0.0726776123046875`)
89
+ - `val_hit_at_20`: `goal ↑ | ↗ +76.67% improving` (early `0.08158111572265625`, mid `0.094146728515625`, late `0.14413070678710938`, last `0.15270233154296875`)
90
+ - `weighted_regularizer_to_contrastive_ratio`: `goal ↓ | ↗ +54.61% improving` (early `0.0667222202939108`, mid `0.0329437448262181`, late `0.03028750823471251`, last `0.028996589606316913`)
91
+ - `samples_per_sec`: `goal ↑ | ↘ -2.88% worsening` (early `1575.5664566260843`, mid `1531.3735887168505`, late `1530.1496274666802`, last `1399.0203442084614`)
92
+ - `gpu_max_memory_gb`: `goal ≈ | ↗ +100.00% improving` (early `13.252925169950275`, mid `13.279059082097401`, late `13.279157161712646`, last `13.279157161712646`)
93
+ - `grad_norm`: `goal ≈ | ↘ -100.00% worsening` (early `4.638080189338309`, mid `4.979553528603791`, late `6.072961742165445`, last `1.7659728527069092`)
94
+
95
+ ## Primary Metric Pressure
96
+
97
+ - `train_loss`: `goal ↓ | ↗ +10.57% improving` (early `6.340343519442343`, mid `5.795929590401622`, late `5.670367905463295`, last `6.491608619689941`)
98
+ - `val_loss`: `goal ↓ | ↘ -55.82% worsening` (early `1.2876316868363775`, mid `2.20697861268016`, late `2.006446816829876`, last `6.646750174145564`)
99
+ - `grad_norm`: `goal ≈ | ↘ -100.00% worsening` (early `4.638080189338309`, mid `4.979553528603791`, late `6.072961742165445`, last `1.7659728527069092`)
100
+ - `contrastive_loss`: `goal ↓ | → -1.64% flat` (early `1.409792113166324`, mid `1.3598130204084982`, late `1.43291388983014`, last `0.5768923759460449`)
101
+ - `structure_bce_loss`: `goal ↓ | ↗ +39.75% improving` (early `0.12293951650489272`, mid `0.07652380578004556`, late `0.07407209518012302`, last `0.08383000642061234`)
102
+ - `val_structure_bce_loss`: `goal ↓ | ↗ +12.20% improving` (early `0.0771060927848642`, mid `0.07029345531468391`, late `0.06769936999660872`, last `0.06532714269997086`)
103
+ - `val_structure_fingerprint_cosine`: `goal ↑ | → +0.17% flat` (early `0.4147872725734487`, mid `0.4204998221830465`, late `0.4154996814322658`, last `0.4255251376889646`)
104
+ - `val_hit_at_20`: `goal ↑ | ↗ +76.67% improving` (early `0.08158111572265625`, mid `0.094146728515625`, late `0.14413070678710938`, last `0.15270233154296875`)
105
+ - `val_structure_top1_top2_gap_mean`: `goal ↑ | ↗ +54.91% improving` (early `0.07839679718017578`, mid `0.1310901641845703`, late `0.12144279479980469`, last `0.11855697631835938`)
106
+ - `val_structure_uncertain_fraction`: `goal ↓ | → -0.03% flat` (early `0.9941177368164062`, mid `0.9762382507324219`, late `0.9944419860839844`, last `0.9922027587890625`)
107
+ - `val_structure_confidence_separation`: `goal ↑ | ↘ -204.85% worsening` (early `-0.00795668972114072`, mid `-0.01309866098972634`, late `-0.024256365849851136`, last `-0.027957430157584895`)
108
+ - `samples_per_sec`: `goal ↑ | ↘ -2.88% worsening` (early `1575.5664566260843`, mid `1531.3735887168505`, late `1530.1496274666802`, last `1399.0203442084614`)
109
+
110
+ ## Debug Metrics
111
+
112
+ - `train_loss`: `6.491608619689941`
113
+ - `train_loss_ema`: `6.519052214060582`
114
+ - `val_loss`: `6.646750174145564`
115
+ - `contrastive_loss`: `0.5768923759460449`
116
+ - `val_ssl_loss`: `1.0397093277424574`
117
+ - `variance_loss`: `0.00213623046875`
118
+ - `covariance_loss`: `10.012618064880371`
119
+ - `structure_bce_loss`: `0.08383000642061234`
120
+ - `structure_retrieval_loss`: `10.630162239074707`
121
+ - `val_structure_retrieval_loss`: `10.882395941764116`
122
+ - `structure_fingerprint_cosine`: `0.46021854877471924`
123
+ - `structure_hit_at_1`: `0.00390625`
124
+ - `structure_hit_at_5`: `0.029296875`
125
+ - `structure_hit_at_20`: `0.1171875`
126
+ - `structure_target_rank_mean`: `314.703125`
127
+ - `structure_logit_margin_mean`: `-1.0078125`
128
+ - `structure_retrieval_logit_std`: `0.9839304089546204`
129
+ - `val_structure_retrieval_logit_std`: `2.6622845763340592`
130
+ - `structure_correct_logit_mean`: `9.1875`
131
+ - `structure_top_negative_logit_mean`: `10.1875`
132
+ - `val_structure_correct_logit_mean`: `9.49462890625`
133
+ - `val_structure_top_negative_logit_mean`: `10.714599609375`
134
+ - `structure_top1_top2_gap_mean`: `0.0986328125`
135
+ - `structure_top1_top2_gap_norm_mean`: `0.10155132412910461`
136
+ - `structure_confidence_candidate_count`: `2126.0`
137
+ - `structure_close_fraction`: `0.029296875`
138
+ - `structure_uncertain_fraction`: `0.970703125`
139
+ - `structure_uncertain_hit_at_1`: `0.0040241447277367115`
140
+ - `structure_confidence_separation`: `-0.0040241447277367115`
141
+ - `val_structure_top1_top2_gap_mean`: `0.11855697631835938`
142
+ - `val_structure_top1_top2_gap_norm_mean`: `0.05411380449368153`
143
+ - `val_structure_confidence_candidate_count`: `4096.0`
144
+ - `val_structure_close_fraction`: `0.0077972412109375`
145
+ - `val_structure_uncertain_fraction`: `0.9922027587890625`
146
+ - `val_structure_close_hit_at_1`: `0.014835291571216658`
147
+ - `val_structure_uncertain_hit_at_1`: `0.027957430157584895`
148
+ - `val_structure_confidence_separation`: `-0.027957430157584895`
149
+ - `grad_norm`: `1.7659728527069092`
150
+ - `weighted_regularizer_to_contrastive_ratio`: `0.028996589606316913`
151
+ - `samples_per_sec`: `1399.0203442084614`
152
+ - `samples_per_sec_60s`: `1399.0203442084614`
153
+ - `samples_per_sec_300s`: `1547.033197547462`
154
+ - `samples_per_sec_900s`: `1512.7885250414204`
155
+ - `samples_per_sec_run_mean`: `1533.0231795943853`
156
+ - `gpu_max_memory_gb`: `13.279157161712646`
157
+ - `gpu_util_mean_pct`: `36.0`
158
+ - `gpu_util_min_pct`: `18.0`
159
+ - `gpu_util_max_pct`: `54.0`
160
+ - `gpu_active_fraction`: `1.0`
161
+ - `gpu_mem_used_mean_gb`: `14.1943359375`
162
+ - `gpu_mem_util_mean_pct`: `14.848754175733243`
163
+ - `gpu_snapshot_available`: `1.0`
164
+ - `step_time_sec`: `0.3659703749981418`
165
+ - `step_time_cv_300s`: `0.06743440395418321`
166
+ - `step_time_cv_900s`: `0.09218780075615579`
167
+ - `dataloader_wait_pct_60s`: `0.0961140629935623`
168
+ - `dataloader_wait_pct_300s`: `0.0382592069912683`
169
+ - `dataloader_wait_pct_900s`: `0.066640439237091`
170
+ - `dataloader_wait_pct_run`: `0.06717802801040498`
171
+ - `mfu_estimate`: `0.06267708747552383`
metadata/checksums.sha256 ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ cef7a30afbc0f4c107dc791039b42edf61bff2337023f9d80b6ff5a2a94bbf29 README.md
2
+ 3c5d7e36bbc4d76a4158562380e32c82248a7b9e4035117d0bbaf900e8184e0c config.json
3
+ 4046e164198deb550b34a4fc54bed0d5b40ccc57a646db5f8c56e261519d8e57 configs/foundation_config.json
4
+ 3b45ebb5770bed57ef75be985fba71513fffdb9ba075d21ab280208e3efa02a4 configs/runtime_requirements.json
5
+ 1debd145d109a2c27d435f8250bcb577366fd143d3977ac48effeeb7674c200f configs/structure_alignment_config.json
6
+ 6a792005b4d2a70b53215b461a36fd9608a39df5898bb1d0a8b35f5af824177c metadata/Final_V3_manifest.json
7
+ e5fb430337eb1516bd168f019e1733b1baf37eb568776e009fb6f348c88e4537 metadata/archive/20260413T201817Z/latest_checkpoint.txt
8
+ e5fb430337eb1516bd168f019e1733b1baf37eb568776e009fb6f348c88e4537 metadata/archive/20260413T202042Z/latest_checkpoint.txt
9
+ a11f8a97e622078c7c0f2fa3f62525f13ed71ff0304deabe84656a7a576855c5 metadata/archive/20260413T213902Z/day1_validation_report.json
10
+ e5fb430337eb1516bd168f019e1733b1baf37eb568776e009fb6f348c88e4537 metadata/archive/20260413T213902Z/latest_checkpoint.txt
11
+ 3e77984c10d6e0f9e8205a142dce9a4c3d513a719b39afda4095b27b92e2cbf2 metadata/archive/20260413T213902Z/latest_metrics.json
12
+ c983893fe5aaf41c63a5d16cf0fdf055196785f23b2cd2ca4cd0a8812fc1ada6 metadata/archive/20260413T213902Z/perf_summary.json
13
+ 2135caf8c7d03112bd8b7fea56da9cf466b1d0f668809d984c33420b179111a8 metadata/archive/20260413T213902Z/results.md
14
+ 631653ad2feae896c11e3ae34f29ec90c317070d5c729bc5178ac3e7d2f23f2d metadata/closeout_receipt.json
15
+ a11f8a97e622078c7c0f2fa3f62525f13ed71ff0304deabe84656a7a576855c5 metadata/day1_validation_report.json
16
+ 6ef7558b9352003473eae0d9b03d9af20b1bd428865378cc5608bfd33cb0d97a metadata/eval_coverage.json
17
+ 0ae4aef28a9733a7bb7f4a2422c09d7dc4f8a8238807be0e5c8e58843c2f3e32 metadata/fast_eval_surface.json
18
+ e5fb430337eb1516bd168f019e1733b1baf37eb568776e009fb6f348c88e4537 metadata/latest_checkpoint.txt
19
+ 3e77984c10d6e0f9e8205a142dce9a4c3d513a719b39afda4095b27b92e2cbf2 metadata/latest_metrics.json
20
+ c983893fe5aaf41c63a5d16cf0fdf055196785f23b2cd2ca4cd0a8812fc1ada6 metadata/perf_summary.json
21
+ 2135caf8c7d03112bd8b7fea56da9cf466b1d0f668809d984c33420b179111a8 metadata/results.md
22
+ d25e392e32ea4aa38534689591582c99ae5d75e007f74390d57d79983adc665a metadata/training_surface.json
23
+ 3a5746d461f17c831bcd08c4180482f7a8424b06e248eaa595202e73929726f0 model_index.json
24
+ d3f6e70ec0b19c0fd8569a4ebd328a117508fe9b333274971967d2da2f9954ff runtime/README.md
25
+ 896284beda3cff6c54e6c3e291ed5c0b2ad8ae708df880831137dee5ed7380db runtime/nexamass_encoder.py
26
+ c16f3bb2ab7c2612b6b84a3359abeaf0a1cb1c055399beac21b39521424328b5 weights/Final_V3-model_state.pt
27
+ 870b30c7bb6a340a460306882fe631d76282f233baf3920ca3a45c363eaf0ee6 weights/Final_V3-model_state.safetensors
28
+ bde396a7448b6972bcb784a465812b4d5b819f8d0c76cf2fba7066db7be9da1f weights/NexaMass-V3-Struct-model_state.pt
29
+ 6e006a6fa4ba6b8f127c7eabc9eaf6b7b64675e0136dee3ca1655bb3a4a3eadd weights/NexaMass-V3-Struct-model_state.safetensors
metadata/closeout_receipt.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_at": "2026-04-13T21:47:04Z",
3
+ "dataset_receipt_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/dataset_receipt.json",
4
+ "dataset_snapshot_uploaded": false,
5
+ "remote_checkpoint_uri": "wasabi:nexa-ms/models/dreams_pretrain_v1/checkpoints",
6
+ "remote_dataset_uri": "wasabi:nexa-ms/processed/gems_v1/receipts",
7
+ "remote_log_uri": "wasabi:nexa-ms/models/dreams_pretrain_v1/logs",
8
+ "run_name": "final_v3_structure_alignment_v26_r1"
9
+ }
metadata/day1_validation_report.json ADDED
@@ -0,0 +1,2475 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "analysis_basis": [
3
+ "docs/projects/msms_pipeline/Technical_Plan.md",
4
+ "docs/projects/msms_pipeline/Engineering.md"
5
+ ],
6
+ "created_at": "2026-04-13T21:36:19Z",
7
+ "dataset_root": "/home/ubuntu/data/gems_v1_handoff_labeled_v26",
8
+ "debug_metric_pressures": {
9
+ "gpu_max_memory_gb": {
10
+ "desired_direction": "stable",
11
+ "early_mean": 13.252925169950275,
12
+ "improvement_pct": 100.0,
13
+ "improvement_pct_clipped": 100.0,
14
+ "last": 13.279157161712646,
15
+ "late_mean": 13.279157161712646,
16
+ "mid_mean": 13.279059082097401,
17
+ "pressure": "pressure_up",
18
+ "pressure_arrow": "\u2197",
19
+ "pressure_score": 1.0
20
+ },
21
+ "structure_confidence_separation": {
22
+ "desired_direction": "higher",
23
+ "early_mean": -0.007918493679557749,
24
+ "improvement_pct": -531.5617836245987,
25
+ "improvement_pct_clipped": -100.0,
26
+ "last": -0.0040241447277367115,
27
+ "late_mean": -0.05001017991881603,
28
+ "mid_mean": -0.032136323077151974,
29
+ "pressure": "pressure_down",
30
+ "pressure_arrow": "\u2198",
31
+ "pressure_score": -1.0
32
+ },
33
+ "structure_correct_logit_mean": {
34
+ "desired_direction": "higher",
35
+ "early_mean": 6.134031791907514,
36
+ "improvement_pct": 47.305071917778086,
37
+ "improvement_pct_clipped": 47.305071917778086,
38
+ "last": 9.1875,
39
+ "late_mean": 9.035739942528735,
40
+ "mid_mean": 9.474891618497109,
41
+ "pressure": "pressure_up",
42
+ "pressure_arrow": "\u2197",
43
+ "pressure_score": 1.0
44
+ },
45
+ "structure_fingerprint_cosine": {
46
+ "desired_direction": "higher",
47
+ "early_mean": 0.3965519580968543,
48
+ "improvement_pct": 17.58747890372286,
49
+ "improvement_pct_clipped": 17.58747890372286,
50
+ "last": 0.46021854877471924,
51
+ "late_mean": 0.46629545006943846,
52
+ "mid_mean": 0.4464927057310336,
53
+ "pressure": "pressure_up",
54
+ "pressure_arrow": "\u2197",
55
+ "pressure_score": 1.0
56
+ },
57
+ "structure_hit_at_1": {
58
+ "desired_direction": "higher",
59
+ "early_mean": 0.007880238439306358,
60
+ "improvement_pct": 549.6830023383724,
61
+ "improvement_pct_clipped": 100.0,
62
+ "last": 0.00390625,
63
+ "late_mean": 0.051196569683908046,
64
+ "mid_mean": 0.032706376445086706,
65
+ "pressure": "pressure_up",
66
+ "pressure_arrow": "\u2197",
67
+ "pressure_score": 1.0
68
+ },
69
+ "structure_hit_at_20": {
70
+ "desired_direction": "higher",
71
+ "early_mean": 0.10721865968208093,
72
+ "improvement_pct": 239.4614028144399,
73
+ "improvement_pct_clipped": 100.0,
74
+ "last": 0.1171875,
75
+ "late_mean": 0.3639659662356322,
76
+ "mid_mean": 0.28105807442196534,
77
+ "pressure": "pressure_up",
78
+ "pressure_arrow": "\u2197",
79
+ "pressure_score": 1.0
80
+ },
81
+ "structure_hit_at_5": {
82
+ "desired_direction": "higher",
83
+ "early_mean": 0.03360955563583815,
84
+ "improvement_pct": 402.47008675709174,
85
+ "improvement_pct_clipped": 100.0,
86
+ "last": 0.029296875,
87
+ "late_mean": 0.16887796336206898,
88
+ "mid_mean": 0.11673591040462428,
89
+ "pressure": "pressure_up",
90
+ "pressure_arrow": "\u2197",
91
+ "pressure_score": 1.0
92
+ },
93
+ "structure_logit_margin_mean": {
94
+ "desired_direction": "higher",
95
+ "early_mean": -1.1263999277456647,
96
+ "improvement_pct": 1.2274364862230593,
97
+ "improvement_pct_clipped": 1.2274364862230593,
98
+ "last": -1.0078125,
99
+ "late_mean": -1.1125740840517242,
100
+ "mid_mean": -1.0154725885115607,
101
+ "pressure": "pressure_none",
102
+ "pressure_arrow": "\u2192",
103
+ "pressure_score": 0.0
104
+ },
105
+ "structure_retrieval_logit_std": {
106
+ "desired_direction": "higher",
107
+ "early_mean": 1.0455416592559375,
108
+ "improvement_pct": 112.05988926255111,
109
+ "improvement_pct_clipped": 100.0,
110
+ "last": 0.9839304089546204,
111
+ "late_mean": 2.2171744848119803,
112
+ "mid_mean": 2.0036371215919537,
113
+ "pressure": "pressure_up",
114
+ "pressure_arrow": "\u2197",
115
+ "pressure_score": 1.0
116
+ },
117
+ "structure_retrieval_loss": {
118
+ "desired_direction": "lower",
119
+ "early_mean": 10.237479978903181,
120
+ "improvement_pct": 10.094066619063353,
121
+ "improvement_pct_clipped": 10.094066619063353,
122
+ "last": 10.630162239074707,
123
+ "late_mean": 9.204101929719421,
124
+ "mid_mean": 9.41523380775672,
125
+ "pressure": "pressure_up",
126
+ "pressure_arrow": "\u2197",
127
+ "pressure_score": 1.0
128
+ },
129
+ "structure_target_rank_mean": {
130
+ "desired_direction": "lower",
131
+ "early_mean": 351.84337743858384,
132
+ "improvement_pct": 58.80171764880722,
133
+ "improvement_pct_clipped": 58.80171764880722,
134
+ "last": 314.703125,
135
+ "late_mean": 144.9534280711207,
136
+ "mid_mean": 176.16389315390174,
137
+ "pressure": "pressure_up",
138
+ "pressure_arrow": "\u2197",
139
+ "pressure_score": 1.0
140
+ },
141
+ "structure_top1_top2_gap_mean": {
142
+ "desired_direction": "higher",
143
+ "early_mean": 0.09793567106213873,
144
+ "improvement_pct": 59.221149173532794,
145
+ "improvement_pct_clipped": 59.221149173532794,
146
+ "last": 0.0986328125,
147
+ "late_mean": 0.1559343009159483,
148
+ "mid_mean": 0.12284365968208093,
149
+ "pressure": "pressure_up",
150
+ "pressure_arrow": "\u2197",
151
+ "pressure_score": 1.0
152
+ },
153
+ "structure_top1_top2_gap_norm_mean": {
154
+ "desired_direction": "higher",
155
+ "early_mean": 0.12215662799301864,
156
+ "improvement_pct": -35.653274040957065,
157
+ "improvement_pct_clipped": -35.653274040957065,
158
+ "last": 0.10155132412910461,
159
+ "late_mean": 0.07860379065547524,
160
+ "mid_mean": 0.07032911504236604,
161
+ "pressure": "pressure_down",
162
+ "pressure_arrow": "\u2198",
163
+ "pressure_score": -1.0
164
+ },
165
+ "structure_top_negative_logit_mean": {
166
+ "desired_direction": "lower",
167
+ "early_mean": 7.262102601156069,
168
+ "improvement_pct": -39.724147080005764,
169
+ "improvement_pct_clipped": -39.724147080005764,
170
+ "last": 10.1875,
171
+ "late_mean": 10.14691091954023,
172
+ "mid_mean": 10.490606936416185,
173
+ "pressure": "pressure_down",
174
+ "pressure_arrow": "\u2198",
175
+ "pressure_score": -1.0
176
+ },
177
+ "structure_uncertain_fraction": {
178
+ "desired_direction": "lower",
179
+ "early_mean": 0.924663565751445,
180
+ "improvement_pct": -6.108024336610137,
181
+ "improvement_pct_clipped": -6.108024336610137,
182
+ "last": 0.970703125,
183
+ "late_mean": 0.9811422413793104,
184
+ "mid_mean": 0.984747561416185,
185
+ "pressure": "pressure_down",
186
+ "pressure_arrow": "\u2198",
187
+ "pressure_score": -1.0
188
+ },
189
+ "train_loss_ema": {
190
+ "desired_direction": "lower",
191
+ "early_mean": 6.265581717362126,
192
+ "improvement_pct": 10.455707205061195,
193
+ "improvement_pct_clipped": 10.455707205061195,
194
+ "last": 6.519052214060582,
195
+ "late_mean": 5.610470838300897,
196
+ "mid_mean": 5.8371378655061745,
197
+ "pressure": "pressure_up",
198
+ "pressure_arrow": "\u2197",
199
+ "pressure_score": 1.0
200
+ },
201
+ "val_ssl_loss": {
202
+ "desired_direction": "lower",
203
+ "early_mean": 1.0849772497183745,
204
+ "improvement_pct": 2.3192742421198114,
205
+ "improvement_pct_clipped": 2.3192742421198114,
206
+ "last": 1.0397093277424574,
207
+ "late_mean": 1.0598136518327963,
208
+ "mid_mean": 1.0747967428483425,
209
+ "pressure": "pressure_none",
210
+ "pressure_arrow": "\u2192",
211
+ "pressure_score": 0.0
212
+ },
213
+ "val_structure_correct_logit_mean": {
214
+ "desired_direction": "higher",
215
+ "early_mean": 8.0311279296875,
216
+ "improvement_pct": 16.44145855816145,
217
+ "improvement_pct_clipped": 16.44145855816145,
218
+ "last": 9.49462890625,
219
+ "late_mean": 9.3515625,
220
+ "mid_mean": 8.441162109375,
221
+ "pressure": "pressure_up",
222
+ "pressure_arrow": "\u2197",
223
+ "pressure_score": 1.0
224
+ },
225
+ "val_structure_retrieval_logit_std": {
226
+ "desired_direction": "higher",
227
+ "early_mean": 1.2326001720502973,
228
+ "improvement_pct": 102.32641409365677,
229
+ "improvement_pct_clipped": 100.0,
230
+ "last": 2.6622845763340592,
231
+ "late_mean": 2.49387572822161,
232
+ "mid_mean": 2.022283688886091,
233
+ "pressure": "pressure_up",
234
+ "pressure_arrow": "\u2197",
235
+ "pressure_score": 1.0
236
+ },
237
+ "val_structure_retrieval_loss": {
238
+ "desired_direction": "lower",
239
+ "early_mean": 11.562767315655947,
240
+ "improvement_pct": 4.864921250164783,
241
+ "improvement_pct_clipped": 4.864921250164783,
242
+ "last": 10.882395941764116,
243
+ "late_mean": 11.000247791409492,
244
+ "mid_mean": 12.150514367967844,
245
+ "pressure": "pressure_none",
246
+ "pressure_arrow": "\u2192",
247
+ "pressure_score": 0.0
248
+ },
249
+ "val_structure_top1_top2_gap_norm_mean": {
250
+ "desired_direction": "higher",
251
+ "early_mean": 0.06950851557485294,
252
+ "improvement_pct": -16.826381575968767,
253
+ "improvement_pct_clipped": -16.826381575968767,
254
+ "last": 0.05411380449368153,
255
+ "late_mean": 0.0578127475164365,
256
+ "mid_mean": 0.07577372609375743,
257
+ "pressure": "pressure_down",
258
+ "pressure_arrow": "\u2198",
259
+ "pressure_score": -1.0
260
+ },
261
+ "val_structure_top_negative_logit_mean": {
262
+ "desired_direction": "lower",
263
+ "early_mean": 9.2255859375,
264
+ "improvement_pct": -16.07653223245475,
265
+ "improvement_pct_clipped": -16.07653223245475,
266
+ "last": 10.714599609375,
267
+ "late_mean": 10.708740234375,
268
+ "mid_mean": 10.51171875,
269
+ "pressure": "pressure_down",
270
+ "pressure_arrow": "\u2198",
271
+ "pressure_score": -1.0
272
+ },
273
+ "weighted_regularizer_to_contrastive_ratio": {
274
+ "desired_direction": "lower",
275
+ "early_mean": 0.0667222202939108,
276
+ "improvement_pct": 54.60656419810926,
277
+ "improvement_pct_clipped": 54.60656419810926,
278
+ "last": 0.028996589606316913,
279
+ "late_mean": 0.03028750823471251,
280
+ "mid_mean": 0.0329437448262181,
281
+ "pressure": "pressure_up",
282
+ "pressure_arrow": "\u2197",
283
+ "pressure_score": 1.0
284
+ }
285
+ },
286
+ "debug_metric_pressures_display": {
287
+ "gpu_max_memory_gb": {
288
+ "desired_direction": "stable",
289
+ "early_mean": 13.252925169950275,
290
+ "goal_display": "goal \u2248",
291
+ "improvement_pct": 100.0,
292
+ "improvement_pct_display": "+100.00%",
293
+ "last": 13.279157161712646,
294
+ "late_mean": 13.279157161712646,
295
+ "mid_mean": 13.279059082097401,
296
+ "pressure": "pressure_up",
297
+ "pressure_arrow": "\u2197",
298
+ "pressure_display": "goal \u2248 | \u2197 +100.00% improving",
299
+ "pressure_score": 1.0,
300
+ "trend_label": "improving"
301
+ },
302
+ "structure_confidence_separation": {
303
+ "desired_direction": "higher",
304
+ "early_mean": -0.007918493679557749,
305
+ "goal_display": "goal \u2191",
306
+ "improvement_pct": -531.5617836245987,
307
+ "improvement_pct_clipped": -100.0,
308
+ "improvement_pct_clipped_display": "-100.00%",
309
+ "improvement_pct_display": "-531.56%",
310
+ "last": -0.0040241447277367115,
311
+ "late_mean": -0.05001017991881603,
312
+ "mid_mean": -0.032136323077151974,
313
+ "pressure": "pressure_down",
314
+ "pressure_arrow": "\u2198",
315
+ "pressure_display": "goal \u2191 | \u2198 -531.56% worsening",
316
+ "pressure_score": -1.0,
317
+ "trend_label": "worsening"
318
+ },
319
+ "structure_correct_logit_mean": {
320
+ "desired_direction": "higher",
321
+ "early_mean": 6.134031791907514,
322
+ "goal_display": "goal \u2191",
323
+ "improvement_pct": 47.305071917778086,
324
+ "improvement_pct_display": "+47.31%",
325
+ "last": 9.1875,
326
+ "late_mean": 9.035739942528735,
327
+ "mid_mean": 9.474891618497109,
328
+ "pressure": "pressure_up",
329
+ "pressure_arrow": "\u2197",
330
+ "pressure_display": "goal \u2191 | \u2197 +47.31% improving",
331
+ "pressure_score": 1.0,
332
+ "trend_label": "improving"
333
+ },
334
+ "structure_fingerprint_cosine": {
335
+ "desired_direction": "higher",
336
+ "early_mean": 0.3965519580968543,
337
+ "goal_display": "goal \u2191",
338
+ "improvement_pct": 17.58747890372286,
339
+ "improvement_pct_display": "+17.59%",
340
+ "last": 0.46021854877471924,
341
+ "late_mean": 0.46629545006943846,
342
+ "mid_mean": 0.4464927057310336,
343
+ "pressure": "pressure_up",
344
+ "pressure_arrow": "\u2197",
345
+ "pressure_display": "goal \u2191 | \u2197 +17.59% improving",
346
+ "pressure_score": 1.0,
347
+ "trend_label": "improving"
348
+ },
349
+ "structure_hit_at_1": {
350
+ "desired_direction": "higher",
351
+ "early_mean": 0.007880238439306358,
352
+ "goal_display": "goal \u2191",
353
+ "improvement_pct": 549.6830023383724,
354
+ "improvement_pct_clipped": 100.0,
355
+ "improvement_pct_clipped_display": "+100.00%",
356
+ "improvement_pct_display": "+549.68%",
357
+ "last": 0.00390625,
358
+ "late_mean": 0.051196569683908046,
359
+ "mid_mean": 0.032706376445086706,
360
+ "pressure": "pressure_up",
361
+ "pressure_arrow": "\u2197",
362
+ "pressure_display": "goal \u2191 | \u2197 +549.68% improving",
363
+ "pressure_score": 1.0,
364
+ "trend_label": "improving"
365
+ },
366
+ "structure_hit_at_20": {
367
+ "desired_direction": "higher",
368
+ "early_mean": 0.10721865968208093,
369
+ "goal_display": "goal \u2191",
370
+ "improvement_pct": 239.4614028144399,
371
+ "improvement_pct_clipped": 100.0,
372
+ "improvement_pct_clipped_display": "+100.00%",
373
+ "improvement_pct_display": "+239.46%",
374
+ "last": 0.1171875,
375
+ "late_mean": 0.3639659662356322,
376
+ "mid_mean": 0.28105807442196534,
377
+ "pressure": "pressure_up",
378
+ "pressure_arrow": "\u2197",
379
+ "pressure_display": "goal \u2191 | \u2197 +239.46% improving",
380
+ "pressure_score": 1.0,
381
+ "trend_label": "improving"
382
+ },
383
+ "structure_hit_at_5": {
384
+ "desired_direction": "higher",
385
+ "early_mean": 0.03360955563583815,
386
+ "goal_display": "goal \u2191",
387
+ "improvement_pct": 402.47008675709174,
388
+ "improvement_pct_clipped": 100.0,
389
+ "improvement_pct_clipped_display": "+100.00%",
390
+ "improvement_pct_display": "+402.47%",
391
+ "last": 0.029296875,
392
+ "late_mean": 0.16887796336206898,
393
+ "mid_mean": 0.11673591040462428,
394
+ "pressure": "pressure_up",
395
+ "pressure_arrow": "\u2197",
396
+ "pressure_display": "goal \u2191 | \u2197 +402.47% improving",
397
+ "pressure_score": 1.0,
398
+ "trend_label": "improving"
399
+ },
400
+ "structure_logit_margin_mean": {
401
+ "desired_direction": "higher",
402
+ "early_mean": -1.1263999277456647,
403
+ "goal_display": "goal \u2191",
404
+ "improvement_pct": 1.2274364862230593,
405
+ "improvement_pct_display": "+1.23%",
406
+ "last": -1.0078125,
407
+ "late_mean": -1.1125740840517242,
408
+ "mid_mean": -1.0154725885115607,
409
+ "pressure": "pressure_none",
410
+ "pressure_arrow": "\u2192",
411
+ "pressure_display": "goal \u2191 | \u2192 +1.23% flat",
412
+ "trend_label": "flat"
413
+ },
414
+ "structure_retrieval_logit_std": {
415
+ "desired_direction": "higher",
416
+ "early_mean": 1.0455416592559375,
417
+ "goal_display": "goal \u2191",
418
+ "improvement_pct": 112.05988926255111,
419
+ "improvement_pct_clipped": 100.0,
420
+ "improvement_pct_clipped_display": "+100.00%",
421
+ "improvement_pct_display": "+112.06%",
422
+ "last": 0.9839304089546204,
423
+ "late_mean": 2.2171744848119803,
424
+ "mid_mean": 2.0036371215919537,
425
+ "pressure": "pressure_up",
426
+ "pressure_arrow": "\u2197",
427
+ "pressure_display": "goal \u2191 | \u2197 +112.06% improving",
428
+ "pressure_score": 1.0,
429
+ "trend_label": "improving"
430
+ },
431
+ "structure_retrieval_loss": {
432
+ "desired_direction": "lower",
433
+ "early_mean": 10.237479978903181,
434
+ "goal_display": "goal \u2193",
435
+ "improvement_pct": 10.094066619063353,
436
+ "improvement_pct_display": "+10.09%",
437
+ "last": 10.630162239074707,
438
+ "late_mean": 9.204101929719421,
439
+ "mid_mean": 9.41523380775672,
440
+ "pressure": "pressure_up",
441
+ "pressure_arrow": "\u2197",
442
+ "pressure_display": "goal \u2193 | \u2197 +10.09% improving",
443
+ "pressure_score": 1.0,
444
+ "trend_label": "improving"
445
+ },
446
+ "structure_target_rank_mean": {
447
+ "desired_direction": "lower",
448
+ "early_mean": 351.84337743858384,
449
+ "goal_display": "goal \u2193",
450
+ "improvement_pct": 58.80171764880722,
451
+ "improvement_pct_display": "+58.80%",
452
+ "last": 314.703125,
453
+ "late_mean": 144.9534280711207,
454
+ "mid_mean": 176.16389315390174,
455
+ "pressure": "pressure_up",
456
+ "pressure_arrow": "\u2197",
457
+ "pressure_display": "goal \u2193 | \u2197 +58.80% improving",
458
+ "pressure_score": 1.0,
459
+ "trend_label": "improving"
460
+ },
461
+ "structure_top1_top2_gap_mean": {
462
+ "desired_direction": "higher",
463
+ "early_mean": 0.09793567106213873,
464
+ "goal_display": "goal \u2191",
465
+ "improvement_pct": 59.221149173532794,
466
+ "improvement_pct_display": "+59.22%",
467
+ "last": 0.0986328125,
468
+ "late_mean": 0.1559343009159483,
469
+ "mid_mean": 0.12284365968208093,
470
+ "pressure": "pressure_up",
471
+ "pressure_arrow": "\u2197",
472
+ "pressure_display": "goal \u2191 | \u2197 +59.22% improving",
473
+ "pressure_score": 1.0,
474
+ "trend_label": "improving"
475
+ },
476
+ "structure_top1_top2_gap_norm_mean": {
477
+ "desired_direction": "higher",
478
+ "early_mean": 0.12215662799301864,
479
+ "goal_display": "goal \u2191",
480
+ "improvement_pct": -35.653274040957065,
481
+ "improvement_pct_display": "-35.65%",
482
+ "last": 0.10155132412910461,
483
+ "late_mean": 0.07860379065547524,
484
+ "mid_mean": 0.07032911504236604,
485
+ "pressure": "pressure_down",
486
+ "pressure_arrow": "\u2198",
487
+ "pressure_display": "goal \u2191 | \u2198 -35.65% worsening",
488
+ "pressure_score": -1.0,
489
+ "trend_label": "worsening"
490
+ },
491
+ "structure_top_negative_logit_mean": {
492
+ "desired_direction": "lower",
493
+ "early_mean": 7.262102601156069,
494
+ "goal_display": "goal \u2193",
495
+ "improvement_pct": -39.724147080005764,
496
+ "improvement_pct_display": "-39.72%",
497
+ "last": 10.1875,
498
+ "late_mean": 10.14691091954023,
499
+ "mid_mean": 10.490606936416185,
500
+ "pressure": "pressure_down",
501
+ "pressure_arrow": "\u2198",
502
+ "pressure_display": "goal \u2193 | \u2198 -39.72% worsening",
503
+ "pressure_score": -1.0,
504
+ "trend_label": "worsening"
505
+ },
506
+ "structure_uncertain_fraction": {
507
+ "desired_direction": "lower",
508
+ "early_mean": 0.924663565751445,
509
+ "goal_display": "goal \u2193",
510
+ "improvement_pct": -6.108024336610137,
511
+ "improvement_pct_display": "-6.11%",
512
+ "last": 0.970703125,
513
+ "late_mean": 0.9811422413793104,
514
+ "mid_mean": 0.984747561416185,
515
+ "pressure": "pressure_down",
516
+ "pressure_arrow": "\u2198",
517
+ "pressure_display": "goal \u2193 | \u2198 -6.11% worsening",
518
+ "pressure_score": -1.0,
519
+ "trend_label": "worsening"
520
+ },
521
+ "train_loss_ema": {
522
+ "desired_direction": "lower",
523
+ "early_mean": 6.265581717362126,
524
+ "goal_display": "goal \u2193",
525
+ "improvement_pct": 10.455707205061195,
526
+ "improvement_pct_display": "+10.46%",
527
+ "last": 6.519052214060582,
528
+ "late_mean": 5.610470838300897,
529
+ "mid_mean": 5.8371378655061745,
530
+ "pressure": "pressure_up",
531
+ "pressure_arrow": "\u2197",
532
+ "pressure_display": "goal \u2193 | \u2197 +10.46% improving",
533
+ "pressure_score": 1.0,
534
+ "trend_label": "improving"
535
+ },
536
+ "val_ssl_loss": {
537
+ "desired_direction": "lower",
538
+ "early_mean": 1.0849772497183745,
539
+ "goal_display": "goal \u2193",
540
+ "improvement_pct": 2.3192742421198114,
541
+ "improvement_pct_display": "+2.32%",
542
+ "last": 1.0397093277424574,
543
+ "late_mean": 1.0598136518327963,
544
+ "mid_mean": 1.0747967428483425,
545
+ "pressure": "pressure_none",
546
+ "pressure_arrow": "\u2192",
547
+ "pressure_display": "goal \u2193 | \u2192 +2.32% flat",
548
+ "trend_label": "flat"
549
+ },
550
+ "val_structure_correct_logit_mean": {
551
+ "desired_direction": "higher",
552
+ "early_mean": 8.0311279296875,
553
+ "goal_display": "goal \u2191",
554
+ "improvement_pct": 16.44145855816145,
555
+ "improvement_pct_display": "+16.44%",
556
+ "last": 9.49462890625,
557
+ "late_mean": 9.3515625,
558
+ "mid_mean": 8.441162109375,
559
+ "pressure": "pressure_up",
560
+ "pressure_arrow": "\u2197",
561
+ "pressure_display": "goal \u2191 | \u2197 +16.44% improving",
562
+ "pressure_score": 1.0,
563
+ "trend_label": "improving"
564
+ },
565
+ "val_structure_retrieval_logit_std": {
566
+ "desired_direction": "higher",
567
+ "early_mean": 1.2326001720502973,
568
+ "goal_display": "goal \u2191",
569
+ "improvement_pct": 102.32641409365677,
570
+ "improvement_pct_clipped": 100.0,
571
+ "improvement_pct_clipped_display": "+100.00%",
572
+ "improvement_pct_display": "+102.33%",
573
+ "last": 2.6622845763340592,
574
+ "late_mean": 2.49387572822161,
575
+ "mid_mean": 2.022283688886091,
576
+ "pressure": "pressure_up",
577
+ "pressure_arrow": "\u2197",
578
+ "pressure_display": "goal \u2191 | \u2197 +102.33% improving",
579
+ "pressure_score": 1.0,
580
+ "trend_label": "improving"
581
+ },
582
+ "val_structure_retrieval_loss": {
583
+ "desired_direction": "lower",
584
+ "early_mean": 11.562767315655947,
585
+ "goal_display": "goal \u2193",
586
+ "improvement_pct": 4.864921250164783,
587
+ "improvement_pct_display": "+4.86%",
588
+ "last": 10.882395941764116,
589
+ "late_mean": 11.000247791409492,
590
+ "mid_mean": 12.150514367967844,
591
+ "pressure": "pressure_none",
592
+ "pressure_arrow": "\u2192",
593
+ "pressure_display": "goal \u2193 | \u2192 +4.86% flat",
594
+ "trend_label": "flat"
595
+ },
596
+ "val_structure_top1_top2_gap_norm_mean": {
597
+ "desired_direction": "higher",
598
+ "early_mean": 0.06950851557485294,
599
+ "goal_display": "goal \u2191",
600
+ "improvement_pct": -16.826381575968767,
601
+ "improvement_pct_display": "-16.83%",
602
+ "last": 0.05411380449368153,
603
+ "late_mean": 0.0578127475164365,
604
+ "mid_mean": 0.07577372609375743,
605
+ "pressure": "pressure_down",
606
+ "pressure_arrow": "\u2198",
607
+ "pressure_display": "goal \u2191 | \u2198 -16.83% worsening",
608
+ "pressure_score": -1.0,
609
+ "trend_label": "worsening"
610
+ },
611
+ "val_structure_top_negative_logit_mean": {
612
+ "desired_direction": "lower",
613
+ "early_mean": 9.2255859375,
614
+ "goal_display": "goal \u2193",
615
+ "improvement_pct": -16.07653223245475,
616
+ "improvement_pct_display": "-16.08%",
617
+ "last": 10.714599609375,
618
+ "late_mean": 10.708740234375,
619
+ "mid_mean": 10.51171875,
620
+ "pressure": "pressure_down",
621
+ "pressure_arrow": "\u2198",
622
+ "pressure_display": "goal \u2193 | \u2198 -16.08% worsening",
623
+ "pressure_score": -1.0,
624
+ "trend_label": "worsening"
625
+ },
626
+ "weighted_regularizer_to_contrastive_ratio": {
627
+ "desired_direction": "lower",
628
+ "early_mean": 0.0667222202939108,
629
+ "goal_display": "goal \u2193",
630
+ "improvement_pct": 54.60656419810926,
631
+ "improvement_pct_display": "+54.61%",
632
+ "last": 0.028996589606316913,
633
+ "late_mean": 0.03028750823471251,
634
+ "mid_mean": 0.0329437448262181,
635
+ "pressure": "pressure_up",
636
+ "pressure_arrow": "\u2197",
637
+ "pressure_display": "goal \u2193 | \u2197 +54.61% improving",
638
+ "pressure_score": 1.0,
639
+ "trend_label": "improving"
640
+ }
641
+ },
642
+ "debug_metrics": {
643
+ "contrastive_loss": 0.5768923759460449,
644
+ "covariance_loss": 10.012618064880371,
645
+ "dataloader_wait_pct_300s": 0.0382592069912683,
646
+ "dataloader_wait_pct_60s": 0.0961140629935623,
647
+ "dataloader_wait_pct_900s": 0.066640439237091,
648
+ "dataloader_wait_pct_run": 0.06717802801040498,
649
+ "gpu_active_fraction": 1.0,
650
+ "gpu_max_memory_gb": 13.279157161712646,
651
+ "gpu_mem_used_mean_gb": 14.1943359375,
652
+ "gpu_mem_util_mean_pct": 14.848754175733243,
653
+ "gpu_snapshot_available": 1.0,
654
+ "gpu_util_max_pct": 54.0,
655
+ "gpu_util_mean_pct": 36.0,
656
+ "gpu_util_min_pct": 18.0,
657
+ "grad_norm": 1.7659728527069092,
658
+ "mfu_estimate": 0.06267708747552383,
659
+ "samples_per_sec": 1399.0203442084614,
660
+ "samples_per_sec_300s": 1547.033197547462,
661
+ "samples_per_sec_60s": 1399.0203442084614,
662
+ "samples_per_sec_900s": 1512.7885250414204,
663
+ "samples_per_sec_run_mean": 1533.0231795943853,
664
+ "step_time_cv_300s": 0.06743440395418321,
665
+ "step_time_cv_900s": 0.09218780075615579,
666
+ "step_time_sec": 0.3659703749981418,
667
+ "structure_bce_loss": 0.08383000642061234,
668
+ "structure_close_fraction": 0.029296875,
669
+ "structure_confidence_candidate_count": 2126.0,
670
+ "structure_confidence_separation": -0.0040241447277367115,
671
+ "structure_correct_logit_mean": 9.1875,
672
+ "structure_fingerprint_cosine": 0.46021854877471924,
673
+ "structure_hit_at_1": 0.00390625,
674
+ "structure_hit_at_20": 0.1171875,
675
+ "structure_hit_at_5": 0.029296875,
676
+ "structure_logit_margin_mean": -1.0078125,
677
+ "structure_retrieval_logit_std": 0.9839304089546204,
678
+ "structure_retrieval_loss": 10.630162239074707,
679
+ "structure_target_rank_mean": 314.703125,
680
+ "structure_top1_top2_gap_mean": 0.0986328125,
681
+ "structure_top1_top2_gap_norm_mean": 0.10155132412910461,
682
+ "structure_top_negative_logit_mean": 10.1875,
683
+ "structure_uncertain_fraction": 0.970703125,
684
+ "structure_uncertain_hit_at_1": 0.0040241447277367115,
685
+ "train_loss": 6.491608619689941,
686
+ "train_loss_ema": 6.519052214060582,
687
+ "val_loss": 6.646750174145564,
688
+ "val_ssl_loss": 1.0397093277424574,
689
+ "val_structure_close_fraction": 0.0077972412109375,
690
+ "val_structure_close_hit_at_1": 0.014835291571216658,
691
+ "val_structure_confidence_candidate_count": 4096.0,
692
+ "val_structure_confidence_separation": -0.027957430157584895,
693
+ "val_structure_correct_logit_mean": 9.49462890625,
694
+ "val_structure_retrieval_logit_std": 2.6622845763340592,
695
+ "val_structure_retrieval_loss": 10.882395941764116,
696
+ "val_structure_top1_top2_gap_mean": 0.11855697631835938,
697
+ "val_structure_top1_top2_gap_norm_mean": 0.05411380449368153,
698
+ "val_structure_top_negative_logit_mean": 10.714599609375,
699
+ "val_structure_uncertain_fraction": 0.9922027587890625,
700
+ "val_structure_uncertain_hit_at_1": 0.027957430157584895,
701
+ "variance_loss": 0.00213623046875,
702
+ "weighted_regularizer_to_contrastive_ratio": 0.028996589606316913
703
+ },
704
+ "eval_coverage": {
705
+ "configured_test_shard_count": 2,
706
+ "configured_test_shards": [
707
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051433_00000.parquet",
708
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051454_00000.parquet"
709
+ ],
710
+ "configured_val_shard_count": 2,
711
+ "configured_val_shards": [
712
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
713
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
714
+ ],
715
+ "created_at": "2026-04-13T21:36:17Z",
716
+ "eval_batch_limit": 256,
717
+ "eval_batches_per_shard_limit": 0,
718
+ "eval_loader_mode": "stream",
719
+ "eval_mode": "deep",
720
+ "full_holdout_coverage": 0.0,
721
+ "full_holdout_sweep": false,
722
+ "full_test_coverage": 0.0,
723
+ "full_val_coverage": 1.0,
724
+ "missing_test_shards": [
725
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051433_00000.parquet",
726
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051454_00000.parquet"
727
+ ],
728
+ "missing_val_shards": [],
729
+ "observed_eval_batches": 256,
730
+ "observed_eval_samples": 131072,
731
+ "observed_test_shard_count": 0,
732
+ "observed_test_shards": [],
733
+ "observed_val_shard_count": 2,
734
+ "observed_val_shards": [
735
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
736
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
737
+ ],
738
+ "per_shard_batch_count": {
739
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet": 128,
740
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet": 128
741
+ },
742
+ "per_shard_sample_count": {
743
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet": 65536,
744
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet": 65536
745
+ },
746
+ "run_step": 10000,
747
+ "step": 91001,
748
+ "unexpected_eval_shards": []
749
+ },
750
+ "failure_modes": [
751
+ {
752
+ "detail": "Observed shard counts {'train': 4, 'val': 2, 'test': 2}; the canonical Day 1 loader contract requires all splits to be readable.",
753
+ "name": "dataset_shards_accessible",
754
+ "status": "pass"
755
+ },
756
+ {
757
+ "detail": "Training batches continued to arrive with positive throughput and no iterator stall.",
758
+ "name": "dataloader_streaming_stable",
759
+ "status": "pass"
760
+ },
761
+ {
762
+ "detail": "Technical_Plan requires zero NaN or Inf tensors and finite train/val loss.",
763
+ "name": "non_finite_loss_or_tensor_failure",
764
+ "status": "pass"
765
+ },
766
+ {
767
+ "detail": "Engineering.md expects slope-aware monitoring; missing validation or retrieval metrics make trend judgment weaker.",
768
+ "name": "validation_observability_gap",
769
+ "status": "pass"
770
+ },
771
+ {
772
+ "detail": "Technical_Plan Day 1 success requires stable GPU memory through the bounded run.",
773
+ "name": "gpu_memory_instability",
774
+ "status": "pass"
775
+ },
776
+ {
777
+ "detail": "Checkpoint save and reload is an explicit Day 1 success criterion.",
778
+ "name": "checkpoint_save_reload_failure",
779
+ "status": "pass"
780
+ },
781
+ {
782
+ "detail": "Engineering.md expects embedding variance monitoring and non-collapsed representations.",
783
+ "name": "embedding_collapse",
784
+ "status": "pass"
785
+ },
786
+ {
787
+ "detail": "The current instability mode is regularization overpowering the SSL objective and producing jagged total loss.",
788
+ "name": "regularizer_dominance",
789
+ "status": "pass"
790
+ },
791
+ {
792
+ "detail": "Once the bridge phase begins, Technical_Plan requires retrieval-oriented validation rather than only augmented contrastive loss.",
793
+ "name": "structure_supervision_gap",
794
+ "status": "pass"
795
+ },
796
+ {
797
+ "detail": "Structure-first qualification now requires explicit confidence diagnostics such as top-1/top-2 separation and uncertainty buckets, not just broader retrieval metrics.",
798
+ "name": "confidence_surface_gap",
799
+ "status": "pass"
800
+ },
801
+ {
802
+ "detail": "The bridge phase should prefer chemistry-grounded fingerprint targets; hashed fallback targets weaken the retrieval signal.",
803
+ "name": "chemistry_target_backend_gap",
804
+ "status": "pass"
805
+ },
806
+ {
807
+ "detail": "Bridge runs must improve retrieval after fingerprint fit saturates; fitting BCE while val Hit@20 stays at candidate-bank baseline means ranking is not generalizing.",
808
+ "name": "bridge_fit_without_ranking",
809
+ "status": "pass"
810
+ },
811
+ {
812
+ "detail": "Engineering.md flags acquisition-artefact shortcut risk; adduct and instrument embeddings remain deliberately down-scaled.",
813
+ "name": "metadata_shortcutting",
814
+ "status": "watch"
815
+ }
816
+ ],
817
+ "fast_eval_surface": {
818
+ "batch_size": 512,
819
+ "batches_per_shard": 0,
820
+ "configured_val_shard_count": 2,
821
+ "configured_val_shards": [
822
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
823
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
824
+ ],
825
+ "created_at": "2026-04-13T20:20:58Z",
826
+ "loader_mode": "stream",
827
+ "per_shard_batch_count": {},
828
+ "per_shard_row_count": {},
829
+ "row_count": 0
830
+ },
831
+ "metric_pressure_summary": {
832
+ "average_improvement_pct": 34.52368967199551,
833
+ "metric_count": 39,
834
+ "overall_arrow": "\u2197",
835
+ "overall_pressure": "pressure_up",
836
+ "pressure_down_count": 11,
837
+ "pressure_none_count": 7,
838
+ "pressure_up_count": 21,
839
+ "pressure_up_fraction": 0.5384615384615384,
840
+ "weighted_improvement_pct": 21.88570985770741,
841
+ "weighted_pressure_score": 0.2838709677419355
842
+ },
843
+ "metric_pressure_summary_display": {
844
+ "average_improvement_pct": 34.52368967199551,
845
+ "average_improvement_pct_display": "+34.52%",
846
+ "metric_count": 39,
847
+ "overall_arrow": "\u2197",
848
+ "overall_display": "\u2197 improving (+21.89% weighted)",
849
+ "overall_label": "improving",
850
+ "overall_pressure": "pressure_up",
851
+ "pressure_down_count": 11,
852
+ "pressure_none_count": 7,
853
+ "pressure_up_count": 21,
854
+ "pressure_up_fraction": 0.5384615384615384,
855
+ "pressure_up_fraction_display": "+53.85%",
856
+ "weighted_improvement_pct": 21.88570985770741,
857
+ "weighted_improvement_pct_display": "+21.89%",
858
+ "weighted_pressure_score": 0.2838709677419355
859
+ },
860
+ "metric_pressures": {
861
+ "contrastive_loss": {
862
+ "desired_direction": "lower",
863
+ "early_mean": 1.409792113166324,
864
+ "improvement_pct": -1.6400841264380104,
865
+ "improvement_pct_clipped": -1.6400841264380104,
866
+ "last": 0.5768923759460449,
867
+ "late_mean": 1.43291388983014,
868
+ "mid_mean": 1.3598130204084982,
869
+ "pressure": "pressure_none",
870
+ "pressure_arrow": "\u2192",
871
+ "pressure_score": 0.0
872
+ },
873
+ "gpu_max_memory_gb": {
874
+ "desired_direction": "stable",
875
+ "early_mean": 13.252925169950275,
876
+ "improvement_pct": 100.0,
877
+ "improvement_pct_clipped": 100.0,
878
+ "last": 13.279157161712646,
879
+ "late_mean": 13.279157161712646,
880
+ "mid_mean": 13.279059082097401,
881
+ "pressure": "pressure_up",
882
+ "pressure_arrow": "\u2197",
883
+ "pressure_score": 1.0
884
+ },
885
+ "grad_norm": {
886
+ "desired_direction": "stable",
887
+ "early_mean": 4.638080189338309,
888
+ "improvement_pct": -100.0,
889
+ "improvement_pct_clipped": -100.0,
890
+ "last": 1.7659728527069092,
891
+ "late_mean": 6.072961742165445,
892
+ "mid_mean": 4.979553528603791,
893
+ "pressure": "pressure_down",
894
+ "pressure_arrow": "\u2198",
895
+ "pressure_score": -1.0
896
+ },
897
+ "samples_per_sec": {
898
+ "desired_direction": "higher",
899
+ "early_mean": 1575.5664566260843,
900
+ "improvement_pct": -2.8825714693532967,
901
+ "improvement_pct_clipped": -2.8825714693532967,
902
+ "last": 1399.0203442084614,
903
+ "late_mean": 1530.1496274666802,
904
+ "mid_mean": 1531.3735887168505,
905
+ "pressure": "pressure_down",
906
+ "pressure_arrow": "\u2198",
907
+ "pressure_score": -1.0
908
+ },
909
+ "structure_bce_loss": {
910
+ "desired_direction": "lower",
911
+ "early_mean": 0.12293951650489272,
912
+ "improvement_pct": 39.749156913940595,
913
+ "improvement_pct_clipped": 39.749156913940595,
914
+ "last": 0.08383000642061234,
915
+ "late_mean": 0.07407209518012302,
916
+ "mid_mean": 0.07652380578004556,
917
+ "pressure": "pressure_up",
918
+ "pressure_arrow": "\u2197",
919
+ "pressure_score": 1.0
920
+ },
921
+ "structure_confidence_separation": {
922
+ "desired_direction": "higher",
923
+ "early_mean": -0.007918493679557749,
924
+ "improvement_pct": -531.5617836245987,
925
+ "improvement_pct_clipped": -100.0,
926
+ "last": -0.0040241447277367115,
927
+ "late_mean": -0.05001017991881603,
928
+ "mid_mean": -0.032136323077151974,
929
+ "pressure": "pressure_down",
930
+ "pressure_arrow": "\u2198",
931
+ "pressure_score": -1.0
932
+ },
933
+ "structure_correct_logit_mean": {
934
+ "desired_direction": "higher",
935
+ "early_mean": 6.134031791907514,
936
+ "improvement_pct": 47.305071917778086,
937
+ "improvement_pct_clipped": 47.305071917778086,
938
+ "last": 9.1875,
939
+ "late_mean": 9.035739942528735,
940
+ "mid_mean": 9.474891618497109,
941
+ "pressure": "pressure_up",
942
+ "pressure_arrow": "\u2197",
943
+ "pressure_score": 1.0
944
+ },
945
+ "structure_fingerprint_cosine": {
946
+ "desired_direction": "higher",
947
+ "early_mean": 0.3965519580968543,
948
+ "improvement_pct": 17.58747890372286,
949
+ "improvement_pct_clipped": 17.58747890372286,
950
+ "last": 0.46021854877471924,
951
+ "late_mean": 0.46629545006943846,
952
+ "mid_mean": 0.4464927057310336,
953
+ "pressure": "pressure_up",
954
+ "pressure_arrow": "\u2197",
955
+ "pressure_score": 1.0
956
+ },
957
+ "structure_hit_at_1": {
958
+ "desired_direction": "higher",
959
+ "early_mean": 0.007880238439306358,
960
+ "improvement_pct": 549.6830023383724,
961
+ "improvement_pct_clipped": 100.0,
962
+ "last": 0.00390625,
963
+ "late_mean": 0.051196569683908046,
964
+ "mid_mean": 0.032706376445086706,
965
+ "pressure": "pressure_up",
966
+ "pressure_arrow": "\u2197",
967
+ "pressure_score": 1.0
968
+ },
969
+ "structure_hit_at_20": {
970
+ "desired_direction": "higher",
971
+ "early_mean": 0.10721865968208093,
972
+ "improvement_pct": 239.4614028144399,
973
+ "improvement_pct_clipped": 100.0,
974
+ "last": 0.1171875,
975
+ "late_mean": 0.3639659662356322,
976
+ "mid_mean": 0.28105807442196534,
977
+ "pressure": "pressure_up",
978
+ "pressure_arrow": "\u2197",
979
+ "pressure_score": 1.0
980
+ },
981
+ "structure_hit_at_5": {
982
+ "desired_direction": "higher",
983
+ "early_mean": 0.03360955563583815,
984
+ "improvement_pct": 402.47008675709174,
985
+ "improvement_pct_clipped": 100.0,
986
+ "last": 0.029296875,
987
+ "late_mean": 0.16887796336206898,
988
+ "mid_mean": 0.11673591040462428,
989
+ "pressure": "pressure_up",
990
+ "pressure_arrow": "\u2197",
991
+ "pressure_score": 1.0
992
+ },
993
+ "structure_logit_margin_mean": {
994
+ "desired_direction": "higher",
995
+ "early_mean": -1.1263999277456647,
996
+ "improvement_pct": 1.2274364862230593,
997
+ "improvement_pct_clipped": 1.2274364862230593,
998
+ "last": -1.0078125,
999
+ "late_mean": -1.1125740840517242,
1000
+ "mid_mean": -1.0154725885115607,
1001
+ "pressure": "pressure_none",
1002
+ "pressure_arrow": "\u2192",
1003
+ "pressure_score": 0.0
1004
+ },
1005
+ "structure_retrieval_logit_std": {
1006
+ "desired_direction": "higher",
1007
+ "early_mean": 1.0455416592559375,
1008
+ "improvement_pct": 112.05988926255111,
1009
+ "improvement_pct_clipped": 100.0,
1010
+ "last": 0.9839304089546204,
1011
+ "late_mean": 2.2171744848119803,
1012
+ "mid_mean": 2.0036371215919537,
1013
+ "pressure": "pressure_up",
1014
+ "pressure_arrow": "\u2197",
1015
+ "pressure_score": 1.0
1016
+ },
1017
+ "structure_retrieval_loss": {
1018
+ "desired_direction": "lower",
1019
+ "early_mean": 10.237479978903181,
1020
+ "improvement_pct": 10.094066619063353,
1021
+ "improvement_pct_clipped": 10.094066619063353,
1022
+ "last": 10.630162239074707,
1023
+ "late_mean": 9.204101929719421,
1024
+ "mid_mean": 9.41523380775672,
1025
+ "pressure": "pressure_up",
1026
+ "pressure_arrow": "\u2197",
1027
+ "pressure_score": 1.0
1028
+ },
1029
+ "structure_target_rank_mean": {
1030
+ "desired_direction": "lower",
1031
+ "early_mean": 351.84337743858384,
1032
+ "improvement_pct": 58.80171764880722,
1033
+ "improvement_pct_clipped": 58.80171764880722,
1034
+ "last": 314.703125,
1035
+ "late_mean": 144.9534280711207,
1036
+ "mid_mean": 176.16389315390174,
1037
+ "pressure": "pressure_up",
1038
+ "pressure_arrow": "\u2197",
1039
+ "pressure_score": 1.0
1040
+ },
1041
+ "structure_top1_top2_gap_mean": {
1042
+ "desired_direction": "higher",
1043
+ "early_mean": 0.09793567106213873,
1044
+ "improvement_pct": 59.221149173532794,
1045
+ "improvement_pct_clipped": 59.221149173532794,
1046
+ "last": 0.0986328125,
1047
+ "late_mean": 0.1559343009159483,
1048
+ "mid_mean": 0.12284365968208093,
1049
+ "pressure": "pressure_up",
1050
+ "pressure_arrow": "\u2197",
1051
+ "pressure_score": 1.0
1052
+ },
1053
+ "structure_top1_top2_gap_norm_mean": {
1054
+ "desired_direction": "higher",
1055
+ "early_mean": 0.12215662799301864,
1056
+ "improvement_pct": -35.653274040957065,
1057
+ "improvement_pct_clipped": -35.653274040957065,
1058
+ "last": 0.10155132412910461,
1059
+ "late_mean": 0.07860379065547524,
1060
+ "mid_mean": 0.07032911504236604,
1061
+ "pressure": "pressure_down",
1062
+ "pressure_arrow": "\u2198",
1063
+ "pressure_score": -1.0
1064
+ },
1065
+ "structure_top_negative_logit_mean": {
1066
+ "desired_direction": "lower",
1067
+ "early_mean": 7.262102601156069,
1068
+ "improvement_pct": -39.724147080005764,
1069
+ "improvement_pct_clipped": -39.724147080005764,
1070
+ "last": 10.1875,
1071
+ "late_mean": 10.14691091954023,
1072
+ "mid_mean": 10.490606936416185,
1073
+ "pressure": "pressure_down",
1074
+ "pressure_arrow": "\u2198",
1075
+ "pressure_score": -1.0
1076
+ },
1077
+ "structure_uncertain_fraction": {
1078
+ "desired_direction": "lower",
1079
+ "early_mean": 0.924663565751445,
1080
+ "improvement_pct": -6.108024336610137,
1081
+ "improvement_pct_clipped": -6.108024336610137,
1082
+ "last": 0.970703125,
1083
+ "late_mean": 0.9811422413793104,
1084
+ "mid_mean": 0.984747561416185,
1085
+ "pressure": "pressure_down",
1086
+ "pressure_arrow": "\u2198",
1087
+ "pressure_score": -1.0
1088
+ },
1089
+ "train_loss": {
1090
+ "desired_direction": "lower",
1091
+ "early_mean": 6.340343519442343,
1092
+ "improvement_pct": 10.566866163080938,
1093
+ "improvement_pct_clipped": 10.566866163080938,
1094
+ "last": 6.491608619689941,
1095
+ "late_mean": 5.670367905463295,
1096
+ "mid_mean": 5.795929590401622,
1097
+ "pressure": "pressure_up",
1098
+ "pressure_arrow": "\u2197",
1099
+ "pressure_score": 1.0
1100
+ },
1101
+ "train_loss_ema": {
1102
+ "desired_direction": "lower",
1103
+ "early_mean": 6.265581717362126,
1104
+ "improvement_pct": 10.455707205061195,
1105
+ "improvement_pct_clipped": 10.455707205061195,
1106
+ "last": 6.519052214060582,
1107
+ "late_mean": 5.610470838300897,
1108
+ "mid_mean": 5.8371378655061745,
1109
+ "pressure": "pressure_up",
1110
+ "pressure_arrow": "\u2197",
1111
+ "pressure_score": 1.0
1112
+ },
1113
+ "val_hit_at_1": {
1114
+ "desired_direction": "higher",
1115
+ "early_mean": 0.0079345703125,
1116
+ "improvement_pct": 207.78846153846155,
1117
+ "improvement_pct_clipped": 100.0,
1118
+ "last": 0.0282135009765625,
1119
+ "late_mean": 0.02442169189453125,
1120
+ "mid_mean": 0.01372528076171875,
1121
+ "pressure": "pressure_none",
1122
+ "pressure_arrow": "\u2192",
1123
+ "pressure_score": 0.0
1124
+ },
1125
+ "val_hit_at_20": {
1126
+ "desired_direction": "higher",
1127
+ "early_mean": 0.08158111572265625,
1128
+ "improvement_pct": 76.67165435331526,
1129
+ "improvement_pct_clipped": 76.67165435331526,
1130
+ "last": 0.15270233154296875,
1131
+ "late_mean": 0.14413070678710938,
1132
+ "mid_mean": 0.094146728515625,
1133
+ "pressure": "pressure_up",
1134
+ "pressure_arrow": "\u2197",
1135
+ "pressure_score": 1.0
1136
+ },
1137
+ "val_hit_at_5": {
1138
+ "desired_direction": "higher",
1139
+ "early_mean": 0.0288543701171875,
1140
+ "improvement_pct": 134.8492860920148,
1141
+ "improvement_pct_clipped": 100.0,
1142
+ "last": 0.0726776123046875,
1143
+ "late_mean": 0.0677642822265625,
1144
+ "mid_mean": 0.0425567626953125,
1145
+ "pressure": "pressure_up",
1146
+ "pressure_arrow": "\u2197",
1147
+ "pressure_score": 1.0
1148
+ },
1149
+ "val_loss": {
1150
+ "desired_direction": "lower",
1151
+ "early_mean": 1.2876316868363775,
1152
+ "improvement_pct": -55.82459156154955,
1153
+ "improvement_pct_clipped": -55.82459156154955,
1154
+ "last": 6.646750174145564,
1155
+ "late_mean": 2.006446816829876,
1156
+ "mid_mean": 2.20697861268016,
1157
+ "pressure": "pressure_down",
1158
+ "pressure_arrow": "\u2198",
1159
+ "pressure_score": -1.0
1160
+ },
1161
+ "val_ssl_loss": {
1162
+ "desired_direction": "lower",
1163
+ "early_mean": 1.0849772497183745,
1164
+ "improvement_pct": 2.3192742421198114,
1165
+ "improvement_pct_clipped": 2.3192742421198114,
1166
+ "last": 1.0397093277424574,
1167
+ "late_mean": 1.0598136518327963,
1168
+ "mid_mean": 1.0747967428483425,
1169
+ "pressure": "pressure_none",
1170
+ "pressure_arrow": "\u2192",
1171
+ "pressure_score": 0.0
1172
+ },
1173
+ "val_structure_bce_loss": {
1174
+ "desired_direction": "lower",
1175
+ "early_mean": 0.0771060927848642,
1176
+ "improvement_pct": 12.199713989530546,
1177
+ "improvement_pct_clipped": 12.199713989530546,
1178
+ "last": 0.06532714269997086,
1179
+ "late_mean": 0.06769936999660872,
1180
+ "mid_mean": 0.07029345531468391,
1181
+ "pressure": "pressure_up",
1182
+ "pressure_arrow": "\u2197",
1183
+ "pressure_score": 1.0
1184
+ },
1185
+ "val_structure_confidence_separation": {
1186
+ "desired_direction": "higher",
1187
+ "early_mean": -0.00795668972114072,
1188
+ "improvement_pct": -204.85499246505236,
1189
+ "improvement_pct_clipped": -100.0,
1190
+ "last": -0.027957430157584895,
1191
+ "late_mean": -0.024256365849851136,
1192
+ "mid_mean": -0.01309866098972634,
1193
+ "pressure": "pressure_down",
1194
+ "pressure_arrow": "\u2198",
1195
+ "pressure_score": -1.0
1196
+ },
1197
+ "val_structure_correct_logit_mean": {
1198
+ "desired_direction": "higher",
1199
+ "early_mean": 8.0311279296875,
1200
+ "improvement_pct": 16.44145855816145,
1201
+ "improvement_pct_clipped": 16.44145855816145,
1202
+ "last": 9.49462890625,
1203
+ "late_mean": 9.3515625,
1204
+ "mid_mean": 8.441162109375,
1205
+ "pressure": "pressure_up",
1206
+ "pressure_arrow": "\u2197",
1207
+ "pressure_score": 1.0
1208
+ },
1209
+ "val_structure_fingerprint_cosine": {
1210
+ "desired_direction": "higher",
1211
+ "early_mean": 0.4147872725734487,
1212
+ "improvement_pct": 0.17175282510408285,
1213
+ "improvement_pct_clipped": 0.17175282510408285,
1214
+ "last": 0.4255251376889646,
1215
+ "late_mean": 0.4154996814322658,
1216
+ "mid_mean": 0.4204998221830465,
1217
+ "pressure": "pressure_none",
1218
+ "pressure_arrow": "\u2192",
1219
+ "pressure_score": 0.0
1220
+ },
1221
+ "val_structure_logit_margin_mean": {
1222
+ "desired_direction": "higher",
1223
+ "early_mean": -1.194091796875,
1224
+ "improvement_pct": -13.669239419341647,
1225
+ "improvement_pct_clipped": -13.669239419341647,
1226
+ "last": -1.221923828125,
1227
+ "late_mean": -1.3573150634765625,
1228
+ "mid_mean": -2.0707931518554688,
1229
+ "pressure": "pressure_down",
1230
+ "pressure_arrow": "\u2198",
1231
+ "pressure_score": -1.0
1232
+ },
1233
+ "val_structure_retrieval_logit_std": {
1234
+ "desired_direction": "higher",
1235
+ "early_mean": 1.2326001720502973,
1236
+ "improvement_pct": 102.32641409365677,
1237
+ "improvement_pct_clipped": 100.0,
1238
+ "last": 2.6622845763340592,
1239
+ "late_mean": 2.49387572822161,
1240
+ "mid_mean": 2.022283688886091,
1241
+ "pressure": "pressure_up",
1242
+ "pressure_arrow": "\u2197",
1243
+ "pressure_score": 1.0
1244
+ },
1245
+ "val_structure_retrieval_loss": {
1246
+ "desired_direction": "lower",
1247
+ "early_mean": 11.562767315655947,
1248
+ "improvement_pct": 4.864921250164783,
1249
+ "improvement_pct_clipped": 4.864921250164783,
1250
+ "last": 10.882395941764116,
1251
+ "late_mean": 11.000247791409492,
1252
+ "mid_mean": 12.150514367967844,
1253
+ "pressure": "pressure_none",
1254
+ "pressure_arrow": "\u2192",
1255
+ "pressure_score": 0.0
1256
+ },
1257
+ "val_structure_target_rank_mean": {
1258
+ "desired_direction": "lower",
1259
+ "early_mean": 760.7540435791016,
1260
+ "improvement_pct": 45.44775026271755,
1261
+ "improvement_pct_clipped": 45.44775026271755,
1262
+ "last": 379.8781280517578,
1263
+ "late_mean": 415.0084457397461,
1264
+ "mid_mean": 683.3385162353516,
1265
+ "pressure": "pressure_up",
1266
+ "pressure_arrow": "\u2197",
1267
+ "pressure_score": 1.0
1268
+ },
1269
+ "val_structure_top1_top2_gap_mean": {
1270
+ "desired_direction": "higher",
1271
+ "early_mean": 0.07839679718017578,
1272
+ "improvement_pct": 54.90785232041847,
1273
+ "improvement_pct_clipped": 54.90785232041847,
1274
+ "last": 0.11855697631835938,
1275
+ "late_mean": 0.12144279479980469,
1276
+ "mid_mean": 0.1310901641845703,
1277
+ "pressure": "pressure_up",
1278
+ "pressure_arrow": "\u2197",
1279
+ "pressure_score": 1.0
1280
+ },
1281
+ "val_structure_top1_top2_gap_norm_mean": {
1282
+ "desired_direction": "higher",
1283
+ "early_mean": 0.06950851557485294,
1284
+ "improvement_pct": -16.826381575968767,
1285
+ "improvement_pct_clipped": -16.826381575968767,
1286
+ "last": 0.05411380449368153,
1287
+ "late_mean": 0.0578127475164365,
1288
+ "mid_mean": 0.07577372609375743,
1289
+ "pressure": "pressure_down",
1290
+ "pressure_arrow": "\u2198",
1291
+ "pressure_score": -1.0
1292
+ },
1293
+ "val_structure_top_negative_logit_mean": {
1294
+ "desired_direction": "lower",
1295
+ "early_mean": 9.2255859375,
1296
+ "improvement_pct": -16.07653223245475,
1297
+ "improvement_pct_clipped": -16.07653223245475,
1298
+ "last": 10.714599609375,
1299
+ "late_mean": 10.708740234375,
1300
+ "mid_mean": 10.51171875,
1301
+ "pressure": "pressure_down",
1302
+ "pressure_arrow": "\u2198",
1303
+ "pressure_score": -1.0
1304
+ },
1305
+ "val_structure_uncertain_fraction": {
1306
+ "desired_direction": "lower",
1307
+ "early_mean": 0.9941177368164062,
1308
+ "improvement_pct": -0.03261678728482514,
1309
+ "improvement_pct_clipped": -0.03261678728482514,
1310
+ "last": 0.9922027587890625,
1311
+ "late_mean": 0.9944419860839844,
1312
+ "mid_mean": 0.9762382507324219,
1313
+ "pressure": "pressure_none",
1314
+ "pressure_arrow": "\u2192",
1315
+ "pressure_score": 0.0
1316
+ },
1317
+ "weighted_regularizer_to_contrastive_ratio": {
1318
+ "desired_direction": "lower",
1319
+ "early_mean": 0.0667222202939108,
1320
+ "improvement_pct": 54.60656419810926,
1321
+ "improvement_pct_clipped": 54.60656419810926,
1322
+ "last": 0.028996589606316913,
1323
+ "late_mean": 0.03028750823471251,
1324
+ "mid_mean": 0.0329437448262181,
1325
+ "pressure": "pressure_up",
1326
+ "pressure_arrow": "\u2197",
1327
+ "pressure_score": 1.0
1328
+ }
1329
+ },
1330
+ "metric_pressures_display": {
1331
+ "contrastive_loss": {
1332
+ "desired_direction": "lower",
1333
+ "early_mean": 1.409792113166324,
1334
+ "goal_display": "goal \u2193",
1335
+ "improvement_pct": -1.6400841264380104,
1336
+ "improvement_pct_display": "-1.64%",
1337
+ "last": 0.5768923759460449,
1338
+ "late_mean": 1.43291388983014,
1339
+ "mid_mean": 1.3598130204084982,
1340
+ "pressure": "pressure_none",
1341
+ "pressure_arrow": "\u2192",
1342
+ "pressure_display": "goal \u2193 | \u2192 -1.64% flat",
1343
+ "trend_label": "flat"
1344
+ },
1345
+ "gpu_max_memory_gb": {
1346
+ "desired_direction": "stable",
1347
+ "early_mean": 13.252925169950275,
1348
+ "goal_display": "goal \u2248",
1349
+ "improvement_pct": 100.0,
1350
+ "improvement_pct_display": "+100.00%",
1351
+ "last": 13.279157161712646,
1352
+ "late_mean": 13.279157161712646,
1353
+ "mid_mean": 13.279059082097401,
1354
+ "pressure": "pressure_up",
1355
+ "pressure_arrow": "\u2197",
1356
+ "pressure_display": "goal \u2248 | \u2197 +100.00% improving",
1357
+ "pressure_score": 1.0,
1358
+ "trend_label": "improving"
1359
+ },
1360
+ "grad_norm": {
1361
+ "desired_direction": "stable",
1362
+ "early_mean": 4.638080189338309,
1363
+ "goal_display": "goal \u2248",
1364
+ "improvement_pct": -100.0,
1365
+ "improvement_pct_display": "-100.00%",
1366
+ "last": 1.7659728527069092,
1367
+ "late_mean": 6.072961742165445,
1368
+ "mid_mean": 4.979553528603791,
1369
+ "pressure": "pressure_down",
1370
+ "pressure_arrow": "\u2198",
1371
+ "pressure_display": "goal \u2248 | \u2198 -100.00% worsening",
1372
+ "pressure_score": -1.0,
1373
+ "trend_label": "worsening"
1374
+ },
1375
+ "samples_per_sec": {
1376
+ "desired_direction": "higher",
1377
+ "early_mean": 1575.5664566260843,
1378
+ "goal_display": "goal \u2191",
1379
+ "improvement_pct": -2.8825714693532967,
1380
+ "improvement_pct_display": "-2.88%",
1381
+ "last": 1399.0203442084614,
1382
+ "late_mean": 1530.1496274666802,
1383
+ "mid_mean": 1531.3735887168505,
1384
+ "pressure": "pressure_down",
1385
+ "pressure_arrow": "\u2198",
1386
+ "pressure_display": "goal \u2191 | \u2198 -2.88% worsening",
1387
+ "pressure_score": -1.0,
1388
+ "trend_label": "worsening"
1389
+ },
1390
+ "structure_bce_loss": {
1391
+ "desired_direction": "lower",
1392
+ "early_mean": 0.12293951650489272,
1393
+ "goal_display": "goal \u2193",
1394
+ "improvement_pct": 39.749156913940595,
1395
+ "improvement_pct_display": "+39.75%",
1396
+ "last": 0.08383000642061234,
1397
+ "late_mean": 0.07407209518012302,
1398
+ "mid_mean": 0.07652380578004556,
1399
+ "pressure": "pressure_up",
1400
+ "pressure_arrow": "\u2197",
1401
+ "pressure_display": "goal \u2193 | \u2197 +39.75% improving",
1402
+ "pressure_score": 1.0,
1403
+ "trend_label": "improving"
1404
+ },
1405
+ "structure_confidence_separation": {
1406
+ "desired_direction": "higher",
1407
+ "early_mean": -0.007918493679557749,
1408
+ "goal_display": "goal \u2191",
1409
+ "improvement_pct": -531.5617836245987,
1410
+ "improvement_pct_clipped": -100.0,
1411
+ "improvement_pct_clipped_display": "-100.00%",
1412
+ "improvement_pct_display": "-531.56%",
1413
+ "last": -0.0040241447277367115,
1414
+ "late_mean": -0.05001017991881603,
1415
+ "mid_mean": -0.032136323077151974,
1416
+ "pressure": "pressure_down",
1417
+ "pressure_arrow": "\u2198",
1418
+ "pressure_display": "goal \u2191 | \u2198 -531.56% worsening",
1419
+ "pressure_score": -1.0,
1420
+ "trend_label": "worsening"
1421
+ },
1422
+ "structure_correct_logit_mean": {
1423
+ "desired_direction": "higher",
1424
+ "early_mean": 6.134031791907514,
1425
+ "goal_display": "goal \u2191",
1426
+ "improvement_pct": 47.305071917778086,
1427
+ "improvement_pct_display": "+47.31%",
1428
+ "last": 9.1875,
1429
+ "late_mean": 9.035739942528735,
1430
+ "mid_mean": 9.474891618497109,
1431
+ "pressure": "pressure_up",
1432
+ "pressure_arrow": "\u2197",
1433
+ "pressure_display": "goal \u2191 | \u2197 +47.31% improving",
1434
+ "pressure_score": 1.0,
1435
+ "trend_label": "improving"
1436
+ },
1437
+ "structure_fingerprint_cosine": {
1438
+ "desired_direction": "higher",
1439
+ "early_mean": 0.3965519580968543,
1440
+ "goal_display": "goal \u2191",
1441
+ "improvement_pct": 17.58747890372286,
1442
+ "improvement_pct_display": "+17.59%",
1443
+ "last": 0.46021854877471924,
1444
+ "late_mean": 0.46629545006943846,
1445
+ "mid_mean": 0.4464927057310336,
1446
+ "pressure": "pressure_up",
1447
+ "pressure_arrow": "\u2197",
1448
+ "pressure_display": "goal \u2191 | \u2197 +17.59% improving",
1449
+ "pressure_score": 1.0,
1450
+ "trend_label": "improving"
1451
+ },
1452
+ "structure_hit_at_1": {
1453
+ "desired_direction": "higher",
1454
+ "early_mean": 0.007880238439306358,
1455
+ "goal_display": "goal \u2191",
1456
+ "improvement_pct": 549.6830023383724,
1457
+ "improvement_pct_clipped": 100.0,
1458
+ "improvement_pct_clipped_display": "+100.00%",
1459
+ "improvement_pct_display": "+549.68%",
1460
+ "last": 0.00390625,
1461
+ "late_mean": 0.051196569683908046,
1462
+ "mid_mean": 0.032706376445086706,
1463
+ "pressure": "pressure_up",
1464
+ "pressure_arrow": "\u2197",
1465
+ "pressure_display": "goal \u2191 | \u2197 +549.68% improving",
1466
+ "pressure_score": 1.0,
1467
+ "trend_label": "improving"
1468
+ },
1469
+ "structure_hit_at_20": {
1470
+ "desired_direction": "higher",
1471
+ "early_mean": 0.10721865968208093,
1472
+ "goal_display": "goal \u2191",
1473
+ "improvement_pct": 239.4614028144399,
1474
+ "improvement_pct_clipped": 100.0,
1475
+ "improvement_pct_clipped_display": "+100.00%",
1476
+ "improvement_pct_display": "+239.46%",
1477
+ "last": 0.1171875,
1478
+ "late_mean": 0.3639659662356322,
1479
+ "mid_mean": 0.28105807442196534,
1480
+ "pressure": "pressure_up",
1481
+ "pressure_arrow": "\u2197",
1482
+ "pressure_display": "goal \u2191 | \u2197 +239.46% improving",
1483
+ "pressure_score": 1.0,
1484
+ "trend_label": "improving"
1485
+ },
1486
+ "structure_hit_at_5": {
1487
+ "desired_direction": "higher",
1488
+ "early_mean": 0.03360955563583815,
1489
+ "goal_display": "goal \u2191",
1490
+ "improvement_pct": 402.47008675709174,
1491
+ "improvement_pct_clipped": 100.0,
1492
+ "improvement_pct_clipped_display": "+100.00%",
1493
+ "improvement_pct_display": "+402.47%",
1494
+ "last": 0.029296875,
1495
+ "late_mean": 0.16887796336206898,
1496
+ "mid_mean": 0.11673591040462428,
1497
+ "pressure": "pressure_up",
1498
+ "pressure_arrow": "\u2197",
1499
+ "pressure_display": "goal \u2191 | \u2197 +402.47% improving",
1500
+ "pressure_score": 1.0,
1501
+ "trend_label": "improving"
1502
+ },
1503
+ "structure_logit_margin_mean": {
1504
+ "desired_direction": "higher",
1505
+ "early_mean": -1.1263999277456647,
1506
+ "goal_display": "goal \u2191",
1507
+ "improvement_pct": 1.2274364862230593,
1508
+ "improvement_pct_display": "+1.23%",
1509
+ "last": -1.0078125,
1510
+ "late_mean": -1.1125740840517242,
1511
+ "mid_mean": -1.0154725885115607,
1512
+ "pressure": "pressure_none",
1513
+ "pressure_arrow": "\u2192",
1514
+ "pressure_display": "goal \u2191 | \u2192 +1.23% flat",
1515
+ "trend_label": "flat"
1516
+ },
1517
+ "structure_retrieval_logit_std": {
1518
+ "desired_direction": "higher",
1519
+ "early_mean": 1.0455416592559375,
1520
+ "goal_display": "goal \u2191",
1521
+ "improvement_pct": 112.05988926255111,
1522
+ "improvement_pct_clipped": 100.0,
1523
+ "improvement_pct_clipped_display": "+100.00%",
1524
+ "improvement_pct_display": "+112.06%",
1525
+ "last": 0.9839304089546204,
1526
+ "late_mean": 2.2171744848119803,
1527
+ "mid_mean": 2.0036371215919537,
1528
+ "pressure": "pressure_up",
1529
+ "pressure_arrow": "\u2197",
1530
+ "pressure_display": "goal \u2191 | \u2197 +112.06% improving",
1531
+ "pressure_score": 1.0,
1532
+ "trend_label": "improving"
1533
+ },
1534
+ "structure_retrieval_loss": {
1535
+ "desired_direction": "lower",
1536
+ "early_mean": 10.237479978903181,
1537
+ "goal_display": "goal \u2193",
1538
+ "improvement_pct": 10.094066619063353,
1539
+ "improvement_pct_display": "+10.09%",
1540
+ "last": 10.630162239074707,
1541
+ "late_mean": 9.204101929719421,
1542
+ "mid_mean": 9.41523380775672,
1543
+ "pressure": "pressure_up",
1544
+ "pressure_arrow": "\u2197",
1545
+ "pressure_display": "goal \u2193 | \u2197 +10.09% improving",
1546
+ "pressure_score": 1.0,
1547
+ "trend_label": "improving"
1548
+ },
1549
+ "structure_target_rank_mean": {
1550
+ "desired_direction": "lower",
1551
+ "early_mean": 351.84337743858384,
1552
+ "goal_display": "goal \u2193",
1553
+ "improvement_pct": 58.80171764880722,
1554
+ "improvement_pct_display": "+58.80%",
1555
+ "last": 314.703125,
1556
+ "late_mean": 144.9534280711207,
1557
+ "mid_mean": 176.16389315390174,
1558
+ "pressure": "pressure_up",
1559
+ "pressure_arrow": "\u2197",
1560
+ "pressure_display": "goal \u2193 | \u2197 +58.80% improving",
1561
+ "pressure_score": 1.0,
1562
+ "trend_label": "improving"
1563
+ },
1564
+ "structure_top1_top2_gap_mean": {
1565
+ "desired_direction": "higher",
1566
+ "early_mean": 0.09793567106213873,
1567
+ "goal_display": "goal \u2191",
1568
+ "improvement_pct": 59.221149173532794,
1569
+ "improvement_pct_display": "+59.22%",
1570
+ "last": 0.0986328125,
1571
+ "late_mean": 0.1559343009159483,
1572
+ "mid_mean": 0.12284365968208093,
1573
+ "pressure": "pressure_up",
1574
+ "pressure_arrow": "\u2197",
1575
+ "pressure_display": "goal \u2191 | \u2197 +59.22% improving",
1576
+ "pressure_score": 1.0,
1577
+ "trend_label": "improving"
1578
+ },
1579
+ "structure_top1_top2_gap_norm_mean": {
1580
+ "desired_direction": "higher",
1581
+ "early_mean": 0.12215662799301864,
1582
+ "goal_display": "goal \u2191",
1583
+ "improvement_pct": -35.653274040957065,
1584
+ "improvement_pct_display": "-35.65%",
1585
+ "last": 0.10155132412910461,
1586
+ "late_mean": 0.07860379065547524,
1587
+ "mid_mean": 0.07032911504236604,
1588
+ "pressure": "pressure_down",
1589
+ "pressure_arrow": "\u2198",
1590
+ "pressure_display": "goal \u2191 | \u2198 -35.65% worsening",
1591
+ "pressure_score": -1.0,
1592
+ "trend_label": "worsening"
1593
+ },
1594
+ "structure_top_negative_logit_mean": {
1595
+ "desired_direction": "lower",
1596
+ "early_mean": 7.262102601156069,
1597
+ "goal_display": "goal \u2193",
1598
+ "improvement_pct": -39.724147080005764,
1599
+ "improvement_pct_display": "-39.72%",
1600
+ "last": 10.1875,
1601
+ "late_mean": 10.14691091954023,
1602
+ "mid_mean": 10.490606936416185,
1603
+ "pressure": "pressure_down",
1604
+ "pressure_arrow": "\u2198",
1605
+ "pressure_display": "goal \u2193 | \u2198 -39.72% worsening",
1606
+ "pressure_score": -1.0,
1607
+ "trend_label": "worsening"
1608
+ },
1609
+ "structure_uncertain_fraction": {
1610
+ "desired_direction": "lower",
1611
+ "early_mean": 0.924663565751445,
1612
+ "goal_display": "goal \u2193",
1613
+ "improvement_pct": -6.108024336610137,
1614
+ "improvement_pct_display": "-6.11%",
1615
+ "last": 0.970703125,
1616
+ "late_mean": 0.9811422413793104,
1617
+ "mid_mean": 0.984747561416185,
1618
+ "pressure": "pressure_down",
1619
+ "pressure_arrow": "\u2198",
1620
+ "pressure_display": "goal \u2193 | \u2198 -6.11% worsening",
1621
+ "pressure_score": -1.0,
1622
+ "trend_label": "worsening"
1623
+ },
1624
+ "train_loss": {
1625
+ "desired_direction": "lower",
1626
+ "early_mean": 6.340343519442343,
1627
+ "goal_display": "goal \u2193",
1628
+ "improvement_pct": 10.566866163080938,
1629
+ "improvement_pct_display": "+10.57%",
1630
+ "last": 6.491608619689941,
1631
+ "late_mean": 5.670367905463295,
1632
+ "mid_mean": 5.795929590401622,
1633
+ "pressure": "pressure_up",
1634
+ "pressure_arrow": "\u2197",
1635
+ "pressure_display": "goal \u2193 | \u2197 +10.57% improving",
1636
+ "pressure_score": 1.0,
1637
+ "trend_label": "improving"
1638
+ },
1639
+ "train_loss_ema": {
1640
+ "desired_direction": "lower",
1641
+ "early_mean": 6.265581717362126,
1642
+ "goal_display": "goal \u2193",
1643
+ "improvement_pct": 10.455707205061195,
1644
+ "improvement_pct_display": "+10.46%",
1645
+ "last": 6.519052214060582,
1646
+ "late_mean": 5.610470838300897,
1647
+ "mid_mean": 5.8371378655061745,
1648
+ "pressure": "pressure_up",
1649
+ "pressure_arrow": "\u2197",
1650
+ "pressure_display": "goal \u2193 | \u2197 +10.46% improving",
1651
+ "pressure_score": 1.0,
1652
+ "trend_label": "improving"
1653
+ },
1654
+ "val_hit_at_1": {
1655
+ "desired_direction": "higher",
1656
+ "early_mean": 0.0079345703125,
1657
+ "goal_display": "goal \u2191",
1658
+ "improvement_pct": 207.78846153846155,
1659
+ "improvement_pct_clipped": 100.0,
1660
+ "improvement_pct_clipped_display": "+100.00%",
1661
+ "improvement_pct_display": "+207.79%",
1662
+ "last": 0.0282135009765625,
1663
+ "late_mean": 0.02442169189453125,
1664
+ "mid_mean": 0.01372528076171875,
1665
+ "pressure": "pressure_none",
1666
+ "pressure_arrow": "\u2192",
1667
+ "pressure_display": "goal \u2191 | \u2192 +207.79% flat",
1668
+ "trend_label": "flat"
1669
+ },
1670
+ "val_hit_at_20": {
1671
+ "desired_direction": "higher",
1672
+ "early_mean": 0.08158111572265625,
1673
+ "goal_display": "goal \u2191",
1674
+ "improvement_pct": 76.67165435331526,
1675
+ "improvement_pct_display": "+76.67%",
1676
+ "last": 0.15270233154296875,
1677
+ "late_mean": 0.14413070678710938,
1678
+ "mid_mean": 0.094146728515625,
1679
+ "pressure": "pressure_up",
1680
+ "pressure_arrow": "\u2197",
1681
+ "pressure_display": "goal \u2191 | \u2197 +76.67% improving",
1682
+ "pressure_score": 1.0,
1683
+ "trend_label": "improving"
1684
+ },
1685
+ "val_hit_at_5": {
1686
+ "desired_direction": "higher",
1687
+ "early_mean": 0.0288543701171875,
1688
+ "goal_display": "goal \u2191",
1689
+ "improvement_pct": 134.8492860920148,
1690
+ "improvement_pct_clipped": 100.0,
1691
+ "improvement_pct_clipped_display": "+100.00%",
1692
+ "improvement_pct_display": "+134.85%",
1693
+ "last": 0.0726776123046875,
1694
+ "late_mean": 0.0677642822265625,
1695
+ "mid_mean": 0.0425567626953125,
1696
+ "pressure": "pressure_up",
1697
+ "pressure_arrow": "\u2197",
1698
+ "pressure_display": "goal \u2191 | \u2197 +134.85% improving",
1699
+ "pressure_score": 1.0,
1700
+ "trend_label": "improving"
1701
+ },
1702
+ "val_loss": {
1703
+ "desired_direction": "lower",
1704
+ "early_mean": 1.2876316868363775,
1705
+ "goal_display": "goal \u2193",
1706
+ "improvement_pct": -55.82459156154955,
1707
+ "improvement_pct_display": "-55.82%",
1708
+ "last": 6.646750174145564,
1709
+ "late_mean": 2.006446816829876,
1710
+ "mid_mean": 2.20697861268016,
1711
+ "pressure": "pressure_down",
1712
+ "pressure_arrow": "\u2198",
1713
+ "pressure_display": "goal \u2193 | \u2198 -55.82% worsening",
1714
+ "pressure_score": -1.0,
1715
+ "trend_label": "worsening"
1716
+ },
1717
+ "val_ssl_loss": {
1718
+ "desired_direction": "lower",
1719
+ "early_mean": 1.0849772497183745,
1720
+ "goal_display": "goal \u2193",
1721
+ "improvement_pct": 2.3192742421198114,
1722
+ "improvement_pct_display": "+2.32%",
1723
+ "last": 1.0397093277424574,
1724
+ "late_mean": 1.0598136518327963,
1725
+ "mid_mean": 1.0747967428483425,
1726
+ "pressure": "pressure_none",
1727
+ "pressure_arrow": "\u2192",
1728
+ "pressure_display": "goal \u2193 | \u2192 +2.32% flat",
1729
+ "trend_label": "flat"
1730
+ },
1731
+ "val_structure_bce_loss": {
1732
+ "desired_direction": "lower",
1733
+ "early_mean": 0.0771060927848642,
1734
+ "goal_display": "goal \u2193",
1735
+ "improvement_pct": 12.199713989530546,
1736
+ "improvement_pct_display": "+12.20%",
1737
+ "last": 0.06532714269997086,
1738
+ "late_mean": 0.06769936999660872,
1739
+ "mid_mean": 0.07029345531468391,
1740
+ "pressure": "pressure_up",
1741
+ "pressure_arrow": "\u2197",
1742
+ "pressure_display": "goal \u2193 | \u2197 +12.20% improving",
1743
+ "pressure_score": 1.0,
1744
+ "trend_label": "improving"
1745
+ },
1746
+ "val_structure_confidence_separation": {
1747
+ "desired_direction": "higher",
1748
+ "early_mean": -0.00795668972114072,
1749
+ "goal_display": "goal \u2191",
1750
+ "improvement_pct": -204.85499246505236,
1751
+ "improvement_pct_clipped": -100.0,
1752
+ "improvement_pct_clipped_display": "-100.00%",
1753
+ "improvement_pct_display": "-204.85%",
1754
+ "last": -0.027957430157584895,
1755
+ "late_mean": -0.024256365849851136,
1756
+ "mid_mean": -0.01309866098972634,
1757
+ "pressure": "pressure_down",
1758
+ "pressure_arrow": "\u2198",
1759
+ "pressure_display": "goal \u2191 | \u2198 -204.85% worsening",
1760
+ "pressure_score": -1.0,
1761
+ "trend_label": "worsening"
1762
+ },
1763
+ "val_structure_correct_logit_mean": {
1764
+ "desired_direction": "higher",
1765
+ "early_mean": 8.0311279296875,
1766
+ "goal_display": "goal \u2191",
1767
+ "improvement_pct": 16.44145855816145,
1768
+ "improvement_pct_display": "+16.44%",
1769
+ "last": 9.49462890625,
1770
+ "late_mean": 9.3515625,
1771
+ "mid_mean": 8.441162109375,
1772
+ "pressure": "pressure_up",
1773
+ "pressure_arrow": "\u2197",
1774
+ "pressure_display": "goal \u2191 | \u2197 +16.44% improving",
1775
+ "pressure_score": 1.0,
1776
+ "trend_label": "improving"
1777
+ },
1778
+ "val_structure_fingerprint_cosine": {
1779
+ "desired_direction": "higher",
1780
+ "early_mean": 0.4147872725734487,
1781
+ "goal_display": "goal \u2191",
1782
+ "improvement_pct": 0.17175282510408285,
1783
+ "improvement_pct_display": "+0.17%",
1784
+ "last": 0.4255251376889646,
1785
+ "late_mean": 0.4154996814322658,
1786
+ "mid_mean": 0.4204998221830465,
1787
+ "pressure": "pressure_none",
1788
+ "pressure_arrow": "\u2192",
1789
+ "pressure_display": "goal \u2191 | \u2192 +0.17% flat",
1790
+ "trend_label": "flat"
1791
+ },
1792
+ "val_structure_logit_margin_mean": {
1793
+ "desired_direction": "higher",
1794
+ "early_mean": -1.194091796875,
1795
+ "goal_display": "goal \u2191",
1796
+ "improvement_pct": -13.669239419341647,
1797
+ "improvement_pct_display": "-13.67%",
1798
+ "last": -1.221923828125,
1799
+ "late_mean": -1.3573150634765625,
1800
+ "mid_mean": -2.0707931518554688,
1801
+ "pressure": "pressure_down",
1802
+ "pressure_arrow": "\u2198",
1803
+ "pressure_display": "goal \u2191 | \u2198 -13.67% worsening",
1804
+ "pressure_score": -1.0,
1805
+ "trend_label": "worsening"
1806
+ },
1807
+ "val_structure_retrieval_logit_std": {
1808
+ "desired_direction": "higher",
1809
+ "early_mean": 1.2326001720502973,
1810
+ "goal_display": "goal \u2191",
1811
+ "improvement_pct": 102.32641409365677,
1812
+ "improvement_pct_clipped": 100.0,
1813
+ "improvement_pct_clipped_display": "+100.00%",
1814
+ "improvement_pct_display": "+102.33%",
1815
+ "last": 2.6622845763340592,
1816
+ "late_mean": 2.49387572822161,
1817
+ "mid_mean": 2.022283688886091,
1818
+ "pressure": "pressure_up",
1819
+ "pressure_arrow": "\u2197",
1820
+ "pressure_display": "goal \u2191 | \u2197 +102.33% improving",
1821
+ "pressure_score": 1.0,
1822
+ "trend_label": "improving"
1823
+ },
1824
+ "val_structure_retrieval_loss": {
1825
+ "desired_direction": "lower",
1826
+ "early_mean": 11.562767315655947,
1827
+ "goal_display": "goal \u2193",
1828
+ "improvement_pct": 4.864921250164783,
1829
+ "improvement_pct_display": "+4.86%",
1830
+ "last": 10.882395941764116,
1831
+ "late_mean": 11.000247791409492,
1832
+ "mid_mean": 12.150514367967844,
1833
+ "pressure": "pressure_none",
1834
+ "pressure_arrow": "\u2192",
1835
+ "pressure_display": "goal \u2193 | \u2192 +4.86% flat",
1836
+ "trend_label": "flat"
1837
+ },
1838
+ "val_structure_target_rank_mean": {
1839
+ "desired_direction": "lower",
1840
+ "early_mean": 760.7540435791016,
1841
+ "goal_display": "goal \u2193",
1842
+ "improvement_pct": 45.44775026271755,
1843
+ "improvement_pct_display": "+45.45%",
1844
+ "last": 379.8781280517578,
1845
+ "late_mean": 415.0084457397461,
1846
+ "mid_mean": 683.3385162353516,
1847
+ "pressure": "pressure_up",
1848
+ "pressure_arrow": "\u2197",
1849
+ "pressure_display": "goal \u2193 | \u2197 +45.45% improving",
1850
+ "pressure_score": 1.0,
1851
+ "trend_label": "improving"
1852
+ },
1853
+ "val_structure_top1_top2_gap_mean": {
1854
+ "desired_direction": "higher",
1855
+ "early_mean": 0.07839679718017578,
1856
+ "goal_display": "goal \u2191",
1857
+ "improvement_pct": 54.90785232041847,
1858
+ "improvement_pct_display": "+54.91%",
1859
+ "last": 0.11855697631835938,
1860
+ "late_mean": 0.12144279479980469,
1861
+ "mid_mean": 0.1310901641845703,
1862
+ "pressure": "pressure_up",
1863
+ "pressure_arrow": "\u2197",
1864
+ "pressure_display": "goal \u2191 | \u2197 +54.91% improving",
1865
+ "pressure_score": 1.0,
1866
+ "trend_label": "improving"
1867
+ },
1868
+ "val_structure_top1_top2_gap_norm_mean": {
1869
+ "desired_direction": "higher",
1870
+ "early_mean": 0.06950851557485294,
1871
+ "goal_display": "goal \u2191",
1872
+ "improvement_pct": -16.826381575968767,
1873
+ "improvement_pct_display": "-16.83%",
1874
+ "last": 0.05411380449368153,
1875
+ "late_mean": 0.0578127475164365,
1876
+ "mid_mean": 0.07577372609375743,
1877
+ "pressure": "pressure_down",
1878
+ "pressure_arrow": "\u2198",
1879
+ "pressure_display": "goal \u2191 | \u2198 -16.83% worsening",
1880
+ "pressure_score": -1.0,
1881
+ "trend_label": "worsening"
1882
+ },
1883
+ "val_structure_top_negative_logit_mean": {
1884
+ "desired_direction": "lower",
1885
+ "early_mean": 9.2255859375,
1886
+ "goal_display": "goal \u2193",
1887
+ "improvement_pct": -16.07653223245475,
1888
+ "improvement_pct_display": "-16.08%",
1889
+ "last": 10.714599609375,
1890
+ "late_mean": 10.708740234375,
1891
+ "mid_mean": 10.51171875,
1892
+ "pressure": "pressure_down",
1893
+ "pressure_arrow": "\u2198",
1894
+ "pressure_display": "goal \u2193 | \u2198 -16.08% worsening",
1895
+ "pressure_score": -1.0,
1896
+ "trend_label": "worsening"
1897
+ },
1898
+ "val_structure_uncertain_fraction": {
1899
+ "desired_direction": "lower",
1900
+ "early_mean": 0.9941177368164062,
1901
+ "goal_display": "goal \u2193",
1902
+ "improvement_pct": -0.03261678728482514,
1903
+ "improvement_pct_display": "-0.03%",
1904
+ "last": 0.9922027587890625,
1905
+ "late_mean": 0.9944419860839844,
1906
+ "mid_mean": 0.9762382507324219,
1907
+ "pressure": "pressure_none",
1908
+ "pressure_arrow": "\u2192",
1909
+ "pressure_display": "goal \u2193 | \u2192 -0.03% flat",
1910
+ "trend_label": "flat"
1911
+ },
1912
+ "weighted_regularizer_to_contrastive_ratio": {
1913
+ "desired_direction": "lower",
1914
+ "early_mean": 0.0667222202939108,
1915
+ "goal_display": "goal \u2193",
1916
+ "improvement_pct": 54.60656419810926,
1917
+ "improvement_pct_display": "+54.61%",
1918
+ "last": 0.028996589606316913,
1919
+ "late_mean": 0.03028750823471251,
1920
+ "mid_mean": 0.0329437448262181,
1921
+ "pressure": "pressure_up",
1922
+ "pressure_arrow": "\u2197",
1923
+ "pressure_display": "goal \u2193 | \u2197 +54.61% improving",
1924
+ "pressure_score": 1.0,
1925
+ "trend_label": "improving"
1926
+ }
1927
+ },
1928
+ "metrics": {
1929
+ "backbone_lr": 3.886130904636371e-06,
1930
+ "checkpoint_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt",
1931
+ "checkpoint_reloaded": 1.0,
1932
+ "contrastive_loss": 0.5768923759460449,
1933
+ "covariance_loss": 10.012618064880371,
1934
+ "covariance_to_contrastive_ratio": 17.356128252623712,
1935
+ "covariance_weight": 0.0015,
1936
+ "dataloader_time_sec": 0.00035174899676349014,
1937
+ "dataloader_wait_pct_300s": 0.0382592069912683,
1938
+ "dataloader_wait_pct_60s": 0.0961140629935623,
1939
+ "dataloader_wait_pct_900s": 0.066640439237091,
1940
+ "dataloader_wait_pct_run": 0.06717802801040498,
1941
+ "different_inchikey_mean_cosine": 0.004679911238535274,
1942
+ "distributed_rank": 0.0,
1943
+ "distributed_world_size": 1.0,
1944
+ "effective_batch_size": 512.0,
1945
+ "effective_hard_negative_scale": 1.0,
1946
+ "effective_retrieval_weight": 0.6,
1947
+ "effective_structure_bce_weight": 1.0,
1948
+ "embedding_mean_norm": 1.0,
1949
+ "embedding_offdiag_mean_cosine": 0.009513536468148232,
1950
+ "embedding_std": 0.07216806709766388,
1951
+ "global_batch_size": 512.0,
1952
+ "gpu_active_fraction": 1.0,
1953
+ "gpu_count_observed": 2.0,
1954
+ "gpu_max_memory_gb": 13.279157161712646,
1955
+ "gpu_mem_used_mean_gb": 14.1943359375,
1956
+ "gpu_mem_util_mean_pct": 14.848754175733243,
1957
+ "gpu_memory_stable": true,
1958
+ "gpu_memory_utilization": 0.13982381154325935,
1959
+ "gpu_snapshot_available": 1.0,
1960
+ "gpu_util_max_pct": 54.0,
1961
+ "gpu_util_mean_pct": 36.0,
1962
+ "gpu_util_min_pct": 18.0,
1963
+ "grad_norm": 1.7659728527069092,
1964
+ "holdout_full_coverage": 0.0,
1965
+ "local_batch_size": 512.0,
1966
+ "lr": 3.886130904636371e-06,
1967
+ "mfu_estimate": 0.06267708747552383,
1968
+ "msms_objective": "hybrid",
1969
+ "optimizer_step": 1.0,
1970
+ "regularizer_ramp": 1.0,
1971
+ "resume_checkpoint_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt",
1972
+ "resume_start_step": 81002.0,
1973
+ "resumed_from_checkpoint": 1.0,
1974
+ "retrieval_head_lr": 1.9430654523181854e-05,
1975
+ "same_inchikey_mean_cosine": 0.16267980930137343,
1976
+ "samples_per_sec": 1399.0203442084614,
1977
+ "samples_per_sec_300s": 1547.033197547462,
1978
+ "samples_per_sec_60s": 1399.0203442084614,
1979
+ "samples_per_sec_900s": 1512.7885250414204,
1980
+ "samples_per_sec_run_mean": 1533.0231795943853,
1981
+ "step_time_cv_300s": 0.06743440395418321,
1982
+ "step_time_cv_60s": 0.0,
1983
+ "step_time_cv_900s": 0.09218780075615579,
1984
+ "step_time_sec": 0.3659703749981418,
1985
+ "structure_bce_locked": 0.0,
1986
+ "structure_bce_loss": 0.08383000642061234,
1987
+ "structure_candidate_count": 2126.0,
1988
+ "structure_clear_fraction": 0.0,
1989
+ "structure_clear_hit_at_1": 0.0,
1990
+ "structure_close_fraction": 0.029296875,
1991
+ "structure_close_hit_at_1": 0.0,
1992
+ "structure_confidence_bucket_score_mean": 0.10155132412910461,
1993
+ "structure_confidence_bucket_score_mode": "normalized_gap",
1994
+ "structure_confidence_candidate_count": 2126.0,
1995
+ "structure_confidence_clear_threshold": 1.0,
1996
+ "structure_confidence_close_threshold": 0.35,
1997
+ "structure_confidence_eval_key_mode": "target_identity",
1998
+ "structure_confidence_separation": -0.0040241447277367115,
1999
+ "structure_correct_logit_mean": 9.1875,
2000
+ "structure_fingerprint_cosine": 0.46021854877471924,
2001
+ "structure_hard_negative_bonus_applied": 1.0,
2002
+ "structure_hard_negative_bonus_mean": 0.060302734375,
2003
+ "structure_hard_negative_scale": 1.0,
2004
+ "structure_head_lr": 1.9430654523181854e-05,
2005
+ "structure_hit_at_1": 0.00390625,
2006
+ "structure_hit_at_20": 0.1171875,
2007
+ "structure_hit_at_5": 0.029296875,
2008
+ "structure_local_ce_loss": 3.545562744140625,
2009
+ "structure_local_ce_loss_weight": 0.4,
2010
+ "structure_local_rerank_loss": 0.0,
2011
+ "structure_local_rerank_loss_weight": 0.0,
2012
+ "structure_local_rerank_score_scale": 1.0,
2013
+ "structure_logit_margin_mean": -1.0078125,
2014
+ "structure_margin_loss": 0.0,
2015
+ "structure_margin_loss_weight": 0.0,
2016
+ "structure_missing_formula_fraction": 0.0,
2017
+ "structure_missing_inchikey_fraction": 0.0,
2018
+ "structure_pairwise_logistic_loss": 0.0,
2019
+ "structure_pairwise_logistic_loss_weight": 0.0,
2020
+ "structure_positive_count_mean": 1.0,
2021
+ "structure_post_label_candidate_count": 2126.0,
2022
+ "structure_post_precursor_candidate_count": 2128.0,
2023
+ "structure_prefilter_candidate_count": 2128.0,
2024
+ "structure_retrieval_logit_std": 0.9839304089546204,
2025
+ "structure_retrieval_loss": 10.630162239074707,
2026
+ "structure_retrieval_scorer": "projected_dot",
2027
+ "structure_retrieval_target_mode": "sample",
2028
+ "structure_skipped_query_fraction": 0.0,
2029
+ "structure_soft_target_exact_mix": 0.4,
2030
+ "structure_soft_target_loss": 7.523618221282959,
2031
+ "structure_soft_target_loss_weight": 0.3,
2032
+ "structure_target_backend": "rdkit_morgan",
2033
+ "structure_target_equivalent_count_mean": 1.0,
2034
+ "structure_target_identity_collision_count": 0.0,
2035
+ "structure_target_rank_mean": 314.703125,
2036
+ "structure_to_ssl_ratio": 217.7124787233957,
2037
+ "structure_top1_top2_gap_mean": 0.0986328125,
2038
+ "structure_top1_top2_gap_norm_mean": 0.10155132412910461,
2039
+ "structure_top_negative_logit_mean": 10.1875,
2040
+ "structure_uncertain_fraction": 0.970703125,
2041
+ "structure_uncertain_hit_at_1": 0.0040241447277367115,
2042
+ "structure_valid_query_fraction": 1.0,
2043
+ "test_expected_shard_count": 2.0,
2044
+ "test_full_coverage": 0.0,
2045
+ "test_observed_shard_count": 0.0,
2046
+ "train_loss": 6.491608619689941,
2047
+ "train_loss_ema": 6.519052214060582,
2048
+ "train_retrieval_bank_size": 4096.0,
2049
+ "val_deep_eval_single_view": 0.0,
2050
+ "val_deep_eval_structure_only": 0.0,
2051
+ "val_effective_structure_bce_weight": 1.0,
2052
+ "val_eval_loader_mode": "stream",
2053
+ "val_eval_mode": "deep",
2054
+ "val_expected_shard_count": 2.0,
2055
+ "val_full_coverage": 1.0,
2056
+ "val_hit_at_1": 0.0282135009765625,
2057
+ "val_hit_at_20": 0.15270233154296875,
2058
+ "val_hit_at_5": 0.0726776123046875,
2059
+ "val_loss": 6.646750174145564,
2060
+ "val_observed_eval_batches": 256.0,
2061
+ "val_observed_shard_count": 2.0,
2062
+ "val_ssl_loss": 1.0397093277424574,
2063
+ "val_structure_bce_loss": 0.06532714269997086,
2064
+ "val_structure_candidate_count": 4119.69140625,
2065
+ "val_structure_clear_fraction": 0.0,
2066
+ "val_structure_clear_hit_at_1": 0.0,
2067
+ "val_structure_close_fraction": 0.0077972412109375,
2068
+ "val_structure_close_hit_at_1": 0.014835291571216658,
2069
+ "val_structure_confidence_candidate_count": 4096.0,
2070
+ "val_structure_confidence_separation": -0.027957430157584895,
2071
+ "val_structure_correct_logit_mean": 9.49462890625,
2072
+ "val_structure_fingerprint_cosine": 0.4255251376889646,
2073
+ "val_structure_logit_margin_mean": -1.221923828125,
2074
+ "val_structure_missing_formula_fraction": 0.0,
2075
+ "val_structure_missing_inchikey_fraction": 0.0,
2076
+ "val_structure_retrieval_logit_std": 2.6622845763340592,
2077
+ "val_structure_retrieval_loss": 10.882395941764116,
2078
+ "val_structure_target_equivalent_count_mean": 1.0,
2079
+ "val_structure_target_identity_collision_count": 23.69140625,
2080
+ "val_structure_target_rank_mean": 379.8781280517578,
2081
+ "val_structure_top1_top2_gap_mean": 0.11855697631835938,
2082
+ "val_structure_top1_top2_gap_norm_mean": 0.05411380449368153,
2083
+ "val_structure_top_negative_logit_mean": 10.714599609375,
2084
+ "val_structure_uncertain_fraction": 0.9922027587890625,
2085
+ "val_structure_uncertain_hit_at_1": 0.027957430157584895,
2086
+ "variance_loss": 0.00213623046875,
2087
+ "variance_weight": 0.8,
2088
+ "weighted_covariance_loss": 0.015018927097320557,
2089
+ "weighted_regularizer_to_contrastive_ratio": 0.028996589606316913,
2090
+ "weighted_ssl_loss": 0.02968101501464844,
2091
+ "weighted_structure_bce_loss": 0.08383000642061234,
2092
+ "weighted_structure_retrieval_loss": 6.378097343444824,
2093
+ "weighted_variance_loss": 0.001708984375
2094
+ },
2095
+ "primary_metric_pressure_summary": {
2096
+ "average_improvement_pct": -14.247321653690681,
2097
+ "metric_count": 12,
2098
+ "overall_arrow": "\u2197",
2099
+ "overall_pressure": "pressure_up",
2100
+ "pressure_down_count": 4,
2101
+ "pressure_none_count": 3,
2102
+ "pressure_up_count": 5,
2103
+ "pressure_up_fraction": 0.4166666666666667,
2104
+ "weighted_improvement_pct": -4.446275812252011,
2105
+ "weighted_pressure_score": 0.1702127659574468
2106
+ },
2107
+ "primary_metric_pressure_summary_display": {
2108
+ "average_improvement_pct": -14.247321653690681,
2109
+ "average_improvement_pct_display": "-14.25%",
2110
+ "metric_count": 12,
2111
+ "overall_arrow": "\u2197",
2112
+ "overall_display": "\u2197 improving (-4.45% weighted)",
2113
+ "overall_label": "improving",
2114
+ "overall_pressure": "pressure_up",
2115
+ "pressure_down_count": 4,
2116
+ "pressure_none_count": 3,
2117
+ "pressure_up_count": 5,
2118
+ "pressure_up_fraction": 0.4166666666666667,
2119
+ "pressure_up_fraction_display": "+41.67%",
2120
+ "weighted_improvement_pct": -4.446275812252011,
2121
+ "weighted_improvement_pct_display": "-4.45%",
2122
+ "weighted_pressure_score": 0.1702127659574468
2123
+ },
2124
+ "primary_metric_pressures": {
2125
+ "contrastive_loss": {
2126
+ "desired_direction": "lower",
2127
+ "early_mean": 1.409792113166324,
2128
+ "improvement_pct": -1.6400841264380104,
2129
+ "improvement_pct_clipped": -1.6400841264380104,
2130
+ "last": 0.5768923759460449,
2131
+ "late_mean": 1.43291388983014,
2132
+ "mid_mean": 1.3598130204084982,
2133
+ "pressure": "pressure_none",
2134
+ "pressure_arrow": "\u2192",
2135
+ "pressure_score": 0.0
2136
+ },
2137
+ "grad_norm": {
2138
+ "desired_direction": "stable",
2139
+ "early_mean": 4.638080189338309,
2140
+ "improvement_pct": -100.0,
2141
+ "improvement_pct_clipped": -100.0,
2142
+ "last": 1.7659728527069092,
2143
+ "late_mean": 6.072961742165445,
2144
+ "mid_mean": 4.979553528603791,
2145
+ "pressure": "pressure_down",
2146
+ "pressure_arrow": "\u2198",
2147
+ "pressure_score": -1.0
2148
+ },
2149
+ "samples_per_sec": {
2150
+ "desired_direction": "higher",
2151
+ "early_mean": 1575.5664566260843,
2152
+ "improvement_pct": -2.8825714693532967,
2153
+ "improvement_pct_clipped": -2.8825714693532967,
2154
+ "last": 1399.0203442084614,
2155
+ "late_mean": 1530.1496274666802,
2156
+ "mid_mean": 1531.3735887168505,
2157
+ "pressure": "pressure_down",
2158
+ "pressure_arrow": "\u2198",
2159
+ "pressure_score": -1.0
2160
+ },
2161
+ "structure_bce_loss": {
2162
+ "desired_direction": "lower",
2163
+ "early_mean": 0.12293951650489272,
2164
+ "improvement_pct": 39.749156913940595,
2165
+ "improvement_pct_clipped": 39.749156913940595,
2166
+ "last": 0.08383000642061234,
2167
+ "late_mean": 0.07407209518012302,
2168
+ "mid_mean": 0.07652380578004556,
2169
+ "pressure": "pressure_up",
2170
+ "pressure_arrow": "\u2197",
2171
+ "pressure_score": 1.0
2172
+ },
2173
+ "train_loss": {
2174
+ "desired_direction": "lower",
2175
+ "early_mean": 6.340343519442343,
2176
+ "improvement_pct": 10.566866163080938,
2177
+ "improvement_pct_clipped": 10.566866163080938,
2178
+ "last": 6.491608619689941,
2179
+ "late_mean": 5.670367905463295,
2180
+ "mid_mean": 5.795929590401622,
2181
+ "pressure": "pressure_up",
2182
+ "pressure_arrow": "\u2197",
2183
+ "pressure_score": 1.0
2184
+ },
2185
+ "val_hit_at_20": {
2186
+ "desired_direction": "higher",
2187
+ "early_mean": 0.08158111572265625,
2188
+ "improvement_pct": 76.67165435331526,
2189
+ "improvement_pct_clipped": 76.67165435331526,
2190
+ "last": 0.15270233154296875,
2191
+ "late_mean": 0.14413070678710938,
2192
+ "mid_mean": 0.094146728515625,
2193
+ "pressure": "pressure_up",
2194
+ "pressure_arrow": "\u2197",
2195
+ "pressure_score": 1.0
2196
+ },
2197
+ "val_loss": {
2198
+ "desired_direction": "lower",
2199
+ "early_mean": 1.2876316868363775,
2200
+ "improvement_pct": -55.82459156154955,
2201
+ "improvement_pct_clipped": -55.82459156154955,
2202
+ "last": 6.646750174145564,
2203
+ "late_mean": 2.006446816829876,
2204
+ "mid_mean": 2.20697861268016,
2205
+ "pressure": "pressure_down",
2206
+ "pressure_arrow": "\u2198",
2207
+ "pressure_score": -1.0
2208
+ },
2209
+ "val_structure_bce_loss": {
2210
+ "desired_direction": "lower",
2211
+ "early_mean": 0.0771060927848642,
2212
+ "improvement_pct": 12.199713989530546,
2213
+ "improvement_pct_clipped": 12.199713989530546,
2214
+ "last": 0.06532714269997086,
2215
+ "late_mean": 0.06769936999660872,
2216
+ "mid_mean": 0.07029345531468391,
2217
+ "pressure": "pressure_up",
2218
+ "pressure_arrow": "\u2197",
2219
+ "pressure_score": 1.0
2220
+ },
2221
+ "val_structure_confidence_separation": {
2222
+ "desired_direction": "higher",
2223
+ "early_mean": -0.00795668972114072,
2224
+ "improvement_pct": -204.85499246505236,
2225
+ "improvement_pct_clipped": -100.0,
2226
+ "last": -0.027957430157584895,
2227
+ "late_mean": -0.024256365849851136,
2228
+ "mid_mean": -0.01309866098972634,
2229
+ "pressure": "pressure_down",
2230
+ "pressure_arrow": "\u2198",
2231
+ "pressure_score": -1.0
2232
+ },
2233
+ "val_structure_fingerprint_cosine": {
2234
+ "desired_direction": "higher",
2235
+ "early_mean": 0.4147872725734487,
2236
+ "improvement_pct": 0.17175282510408285,
2237
+ "improvement_pct_clipped": 0.17175282510408285,
2238
+ "last": 0.4255251376889646,
2239
+ "late_mean": 0.4154996814322658,
2240
+ "mid_mean": 0.4204998221830465,
2241
+ "pressure": "pressure_none",
2242
+ "pressure_arrow": "\u2192",
2243
+ "pressure_score": 0.0
2244
+ },
2245
+ "val_structure_top1_top2_gap_mean": {
2246
+ "desired_direction": "higher",
2247
+ "early_mean": 0.07839679718017578,
2248
+ "improvement_pct": 54.90785232041847,
2249
+ "improvement_pct_clipped": 54.90785232041847,
2250
+ "last": 0.11855697631835938,
2251
+ "late_mean": 0.12144279479980469,
2252
+ "mid_mean": 0.1310901641845703,
2253
+ "pressure": "pressure_up",
2254
+ "pressure_arrow": "\u2197",
2255
+ "pressure_score": 1.0
2256
+ },
2257
+ "val_structure_uncertain_fraction": {
2258
+ "desired_direction": "lower",
2259
+ "early_mean": 0.9941177368164062,
2260
+ "improvement_pct": -0.03261678728482514,
2261
+ "improvement_pct_clipped": -0.03261678728482514,
2262
+ "last": 0.9922027587890625,
2263
+ "late_mean": 0.9944419860839844,
2264
+ "mid_mean": 0.9762382507324219,
2265
+ "pressure": "pressure_none",
2266
+ "pressure_arrow": "\u2192",
2267
+ "pressure_score": 0.0
2268
+ }
2269
+ },
2270
+ "primary_metric_pressures_display": {
2271
+ "contrastive_loss": {
2272
+ "desired_direction": "lower",
2273
+ "early_mean": 1.409792113166324,
2274
+ "goal_display": "goal \u2193",
2275
+ "improvement_pct": -1.6400841264380104,
2276
+ "improvement_pct_display": "-1.64%",
2277
+ "last": 0.5768923759460449,
2278
+ "late_mean": 1.43291388983014,
2279
+ "mid_mean": 1.3598130204084982,
2280
+ "pressure": "pressure_none",
2281
+ "pressure_arrow": "\u2192",
2282
+ "pressure_display": "goal \u2193 | \u2192 -1.64% flat",
2283
+ "trend_label": "flat"
2284
+ },
2285
+ "grad_norm": {
2286
+ "desired_direction": "stable",
2287
+ "early_mean": 4.638080189338309,
2288
+ "goal_display": "goal \u2248",
2289
+ "improvement_pct": -100.0,
2290
+ "improvement_pct_display": "-100.00%",
2291
+ "last": 1.7659728527069092,
2292
+ "late_mean": 6.072961742165445,
2293
+ "mid_mean": 4.979553528603791,
2294
+ "pressure": "pressure_down",
2295
+ "pressure_arrow": "\u2198",
2296
+ "pressure_display": "goal \u2248 | \u2198 -100.00% worsening",
2297
+ "pressure_score": -1.0,
2298
+ "trend_label": "worsening"
2299
+ },
2300
+ "samples_per_sec": {
2301
+ "desired_direction": "higher",
2302
+ "early_mean": 1575.5664566260843,
2303
+ "goal_display": "goal \u2191",
2304
+ "improvement_pct": -2.8825714693532967,
2305
+ "improvement_pct_display": "-2.88%",
2306
+ "last": 1399.0203442084614,
2307
+ "late_mean": 1530.1496274666802,
2308
+ "mid_mean": 1531.3735887168505,
2309
+ "pressure": "pressure_down",
2310
+ "pressure_arrow": "\u2198",
2311
+ "pressure_display": "goal \u2191 | \u2198 -2.88% worsening",
2312
+ "pressure_score": -1.0,
2313
+ "trend_label": "worsening"
2314
+ },
2315
+ "structure_bce_loss": {
2316
+ "desired_direction": "lower",
2317
+ "early_mean": 0.12293951650489272,
2318
+ "goal_display": "goal \u2193",
2319
+ "improvement_pct": 39.749156913940595,
2320
+ "improvement_pct_display": "+39.75%",
2321
+ "last": 0.08383000642061234,
2322
+ "late_mean": 0.07407209518012302,
2323
+ "mid_mean": 0.07652380578004556,
2324
+ "pressure": "pressure_up",
2325
+ "pressure_arrow": "\u2197",
2326
+ "pressure_display": "goal \u2193 | \u2197 +39.75% improving",
2327
+ "pressure_score": 1.0,
2328
+ "trend_label": "improving"
2329
+ },
2330
+ "train_loss": {
2331
+ "desired_direction": "lower",
2332
+ "early_mean": 6.340343519442343,
2333
+ "goal_display": "goal \u2193",
2334
+ "improvement_pct": 10.566866163080938,
2335
+ "improvement_pct_display": "+10.57%",
2336
+ "last": 6.491608619689941,
2337
+ "late_mean": 5.670367905463295,
2338
+ "mid_mean": 5.795929590401622,
2339
+ "pressure": "pressure_up",
2340
+ "pressure_arrow": "\u2197",
2341
+ "pressure_display": "goal \u2193 | \u2197 +10.57% improving",
2342
+ "pressure_score": 1.0,
2343
+ "trend_label": "improving"
2344
+ },
2345
+ "val_hit_at_20": {
2346
+ "desired_direction": "higher",
2347
+ "early_mean": 0.08158111572265625,
2348
+ "goal_display": "goal \u2191",
2349
+ "improvement_pct": 76.67165435331526,
2350
+ "improvement_pct_display": "+76.67%",
2351
+ "last": 0.15270233154296875,
2352
+ "late_mean": 0.14413070678710938,
2353
+ "mid_mean": 0.094146728515625,
2354
+ "pressure": "pressure_up",
2355
+ "pressure_arrow": "\u2197",
2356
+ "pressure_display": "goal \u2191 | \u2197 +76.67% improving",
2357
+ "pressure_score": 1.0,
2358
+ "trend_label": "improving"
2359
+ },
2360
+ "val_loss": {
2361
+ "desired_direction": "lower",
2362
+ "early_mean": 1.2876316868363775,
2363
+ "goal_display": "goal \u2193",
2364
+ "improvement_pct": -55.82459156154955,
2365
+ "improvement_pct_display": "-55.82%",
2366
+ "last": 6.646750174145564,
2367
+ "late_mean": 2.006446816829876,
2368
+ "mid_mean": 2.20697861268016,
2369
+ "pressure": "pressure_down",
2370
+ "pressure_arrow": "\u2198",
2371
+ "pressure_display": "goal \u2193 | \u2198 -55.82% worsening",
2372
+ "pressure_score": -1.0,
2373
+ "trend_label": "worsening"
2374
+ },
2375
+ "val_structure_bce_loss": {
2376
+ "desired_direction": "lower",
2377
+ "early_mean": 0.0771060927848642,
2378
+ "goal_display": "goal \u2193",
2379
+ "improvement_pct": 12.199713989530546,
2380
+ "improvement_pct_display": "+12.20%",
2381
+ "last": 0.06532714269997086,
2382
+ "late_mean": 0.06769936999660872,
2383
+ "mid_mean": 0.07029345531468391,
2384
+ "pressure": "pressure_up",
2385
+ "pressure_arrow": "\u2197",
2386
+ "pressure_display": "goal \u2193 | \u2197 +12.20% improving",
2387
+ "pressure_score": 1.0,
2388
+ "trend_label": "improving"
2389
+ },
2390
+ "val_structure_confidence_separation": {
2391
+ "desired_direction": "higher",
2392
+ "early_mean": -0.00795668972114072,
2393
+ "goal_display": "goal \u2191",
2394
+ "improvement_pct": -204.85499246505236,
2395
+ "improvement_pct_clipped": -100.0,
2396
+ "improvement_pct_clipped_display": "-100.00%",
2397
+ "improvement_pct_display": "-204.85%",
2398
+ "last": -0.027957430157584895,
2399
+ "late_mean": -0.024256365849851136,
2400
+ "mid_mean": -0.01309866098972634,
2401
+ "pressure": "pressure_down",
2402
+ "pressure_arrow": "\u2198",
2403
+ "pressure_display": "goal \u2191 | \u2198 -204.85% worsening",
2404
+ "pressure_score": -1.0,
2405
+ "trend_label": "worsening"
2406
+ },
2407
+ "val_structure_fingerprint_cosine": {
2408
+ "desired_direction": "higher",
2409
+ "early_mean": 0.4147872725734487,
2410
+ "goal_display": "goal \u2191",
2411
+ "improvement_pct": 0.17175282510408285,
2412
+ "improvement_pct_display": "+0.17%",
2413
+ "last": 0.4255251376889646,
2414
+ "late_mean": 0.4154996814322658,
2415
+ "mid_mean": 0.4204998221830465,
2416
+ "pressure": "pressure_none",
2417
+ "pressure_arrow": "\u2192",
2418
+ "pressure_display": "goal \u2191 | \u2192 +0.17% flat",
2419
+ "trend_label": "flat"
2420
+ },
2421
+ "val_structure_top1_top2_gap_mean": {
2422
+ "desired_direction": "higher",
2423
+ "early_mean": 0.07839679718017578,
2424
+ "goal_display": "goal \u2191",
2425
+ "improvement_pct": 54.90785232041847,
2426
+ "improvement_pct_display": "+54.91%",
2427
+ "last": 0.11855697631835938,
2428
+ "late_mean": 0.12144279479980469,
2429
+ "mid_mean": 0.1310901641845703,
2430
+ "pressure": "pressure_up",
2431
+ "pressure_arrow": "\u2197",
2432
+ "pressure_display": "goal \u2191 | \u2197 +54.91% improving",
2433
+ "pressure_score": 1.0,
2434
+ "trend_label": "improving"
2435
+ },
2436
+ "val_structure_uncertain_fraction": {
2437
+ "desired_direction": "lower",
2438
+ "early_mean": 0.9941177368164062,
2439
+ "goal_display": "goal \u2193",
2440
+ "improvement_pct": -0.03261678728482514,
2441
+ "improvement_pct_display": "-0.03%",
2442
+ "last": 0.9922027587890625,
2443
+ "late_mean": 0.9944419860839844,
2444
+ "mid_mean": 0.9762382507324219,
2445
+ "pressure": "pressure_none",
2446
+ "pressure_arrow": "\u2192",
2447
+ "pressure_display": "goal \u2193 | \u2192 -0.03% flat",
2448
+ "trend_label": "flat"
2449
+ }
2450
+ },
2451
+ "primary_metrics": {
2452
+ "contrastive_loss": 0.5768923759460449,
2453
+ "different_inchikey_mean_cosine": 0.004679911238535274,
2454
+ "embedding_std": 0.07216806709766388,
2455
+ "grad_norm": 1.7659728527069092,
2456
+ "samples_per_sec": 1399.0203442084614,
2457
+ "structure_bce_loss": 0.08383000642061234,
2458
+ "train_loss": 6.491608619689941,
2459
+ "val_hit_at_20": 0.15270233154296875,
2460
+ "val_loss": 6.646750174145564,
2461
+ "val_structure_bce_loss": 0.06532714269997086,
2462
+ "val_structure_candidate_count": 4119.69140625,
2463
+ "val_structure_clear_hit_at_1": 0.0,
2464
+ "val_structure_confidence_separation": -0.027957430157584895,
2465
+ "val_structure_fingerprint_cosine": 0.4255251376889646,
2466
+ "val_structure_top1_top2_gap_mean": 0.11855697631835938,
2467
+ "val_structure_uncertain_fraction": 0.9922027587890625
2468
+ },
2469
+ "run_name": "final_v3_structure_alignment_v26_r1",
2470
+ "shard_counts": {
2471
+ "test": 2,
2472
+ "train": 4,
2473
+ "val": 2
2474
+ }
2475
+ }
metadata/eval_coverage.json ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_at": "2026-04-13T21:36:17Z",
3
+ "step": 91001,
4
+ "run_step": 10000,
5
+ "eval_mode": "deep",
6
+ "eval_loader_mode": "stream",
7
+ "eval_batch_limit": 256,
8
+ "eval_batches_per_shard_limit": 0,
9
+ "full_holdout_sweep": false,
10
+ "configured_val_shard_count": 2,
11
+ "configured_val_shards": [
12
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
13
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
14
+ ],
15
+ "configured_test_shard_count": 2,
16
+ "configured_test_shards": [
17
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051433_00000.parquet",
18
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051454_00000.parquet"
19
+ ],
20
+ "observed_val_shard_count": 2,
21
+ "observed_val_shards": [
22
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
23
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
24
+ ],
25
+ "missing_val_shards": [],
26
+ "observed_test_shard_count": 0,
27
+ "observed_test_shards": [],
28
+ "missing_test_shards": [
29
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051433_00000.parquet",
30
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/test/shard_20260313_051454_00000.parquet"
31
+ ],
32
+ "unexpected_eval_shards": [],
33
+ "observed_eval_batches": 256,
34
+ "observed_eval_samples": 131072,
35
+ "per_shard_batch_count": {
36
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet": 128,
37
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet": 128
38
+ },
39
+ "per_shard_sample_count": {
40
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet": 65536,
41
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet": 65536
42
+ },
43
+ "full_val_coverage": 1.0,
44
+ "full_test_coverage": 0.0,
45
+ "full_holdout_coverage": 0.0
46
+ }
metadata/fast_eval_surface.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_at": "2026-04-13T20:20:58Z",
3
+ "loader_mode": "stream",
4
+ "configured_val_shard_count": 2,
5
+ "configured_val_shards": [
6
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051433_00000.parquet",
7
+ "/home/ubuntu/data/gems_v1_handoff_labeled_v26/shards/val/shard_20260313_051454_00000.parquet"
8
+ ],
9
+ "batches_per_shard": 0,
10
+ "batch_size": 512,
11
+ "row_count": 0,
12
+ "per_shard_row_count": {},
13
+ "per_shard_batch_count": {}
14
+ }
metadata/latest_checkpoint.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ /home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt
metadata/latest_metrics.json ADDED
@@ -0,0 +1,169 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backbone_lr": 3.886130904636371e-06,
3
+ "checkpoint_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt",
4
+ "checkpoint_reloaded": 1.0,
5
+ "contrastive_loss": 0.5768923759460449,
6
+ "covariance_loss": 10.012618064880371,
7
+ "covariance_to_contrastive_ratio": 17.356128252623712,
8
+ "covariance_weight": 0.0015,
9
+ "dataloader_time_sec": 0.00035174899676349014,
10
+ "dataloader_wait_pct_300s": 0.0382592069912683,
11
+ "dataloader_wait_pct_60s": 0.0961140629935623,
12
+ "dataloader_wait_pct_900s": 0.066640439237091,
13
+ "dataloader_wait_pct_run": 0.06717802801040498,
14
+ "different_inchikey_mean_cosine": 0.004679911238535274,
15
+ "distributed_rank": 0.0,
16
+ "distributed_world_size": 1.0,
17
+ "effective_batch_size": 512.0,
18
+ "effective_hard_negative_scale": 1.0,
19
+ "effective_retrieval_weight": 0.6,
20
+ "effective_structure_bce_weight": 1.0,
21
+ "embedding_mean_norm": 1.0,
22
+ "embedding_offdiag_mean_cosine": 0.009513536468148232,
23
+ "embedding_std": 0.07216806709766388,
24
+ "final_report": 1.0,
25
+ "global_batch_size": 512.0,
26
+ "gpu_active_fraction": 1.0,
27
+ "gpu_count_observed": 2.0,
28
+ "gpu_max_memory_gb": 13.279157161712646,
29
+ "gpu_mem_used_mean_gb": 14.1943359375,
30
+ "gpu_mem_util_mean_pct": 14.848754175733243,
31
+ "gpu_memory_stable": true,
32
+ "gpu_memory_utilization": 0.13982381154325935,
33
+ "gpu_snapshot_available": 1.0,
34
+ "gpu_util_max_pct": 54.0,
35
+ "gpu_util_mean_pct": 36.0,
36
+ "gpu_util_min_pct": 18.0,
37
+ "grad_norm": 1.7659728527069092,
38
+ "holdout_full_coverage": 0.0,
39
+ "local_batch_size": 512.0,
40
+ "lr": 3.886130904636371e-06,
41
+ "mfu_estimate": 0.06267708747552383,
42
+ "msms_objective": "hybrid",
43
+ "optimizer_step": 1.0,
44
+ "regularizer_ramp": 1.0,
45
+ "resume_checkpoint_path": "/home/ubuntu/artifacts/nexa_mol/final_v3_structure_alignment/v26_r1/current_checkpoint.pt",
46
+ "resume_start_step": 81002.0,
47
+ "resumed_from_checkpoint": 1.0,
48
+ "retrieval_head_lr": 1.9430654523181854e-05,
49
+ "same_inchikey_mean_cosine": 0.16267980930137343,
50
+ "samples_per_sec": 1399.0203442084614,
51
+ "samples_per_sec_300s": 1547.033197547462,
52
+ "samples_per_sec_60s": 1399.0203442084614,
53
+ "samples_per_sec_900s": 1512.7885250414204,
54
+ "samples_per_sec_run_mean": 1533.0231795943853,
55
+ "step": 91001,
56
+ "step_time_cv_300s": 0.06743440395418321,
57
+ "step_time_cv_60s": 0.0,
58
+ "step_time_cv_900s": 0.09218780075615579,
59
+ "step_time_sec": 0.3659703749981418,
60
+ "structure_bce_locked": 0.0,
61
+ "structure_bce_loss": 0.08383000642061234,
62
+ "structure_candidate_count": 2126.0,
63
+ "structure_clear_fraction": 0.0,
64
+ "structure_clear_hit_at_1": 0.0,
65
+ "structure_close_fraction": 0.029296875,
66
+ "structure_close_hit_at_1": 0.0,
67
+ "structure_confidence_bucket_score_mean": 0.10155132412910461,
68
+ "structure_confidence_bucket_score_mode": "normalized_gap",
69
+ "structure_confidence_candidate_count": 2126.0,
70
+ "structure_confidence_clear_threshold": 1.0,
71
+ "structure_confidence_close_threshold": 0.35,
72
+ "structure_confidence_eval_key_mode": "target_identity",
73
+ "structure_confidence_separation": -0.0040241447277367115,
74
+ "structure_correct_logit_mean": 9.1875,
75
+ "structure_fingerprint_cosine": 0.46021854877471924,
76
+ "structure_hard_negative_bonus_applied": 1.0,
77
+ "structure_hard_negative_bonus_mean": 0.060302734375,
78
+ "structure_hard_negative_scale": 1.0,
79
+ "structure_head_lr": 1.9430654523181854e-05,
80
+ "structure_hit_at_1": 0.00390625,
81
+ "structure_hit_at_20": 0.1171875,
82
+ "structure_hit_at_5": 0.029296875,
83
+ "structure_local_ce_loss": 3.545562744140625,
84
+ "structure_local_ce_loss_weight": 0.4,
85
+ "structure_local_rerank_loss": 0.0,
86
+ "structure_local_rerank_loss_weight": 0.0,
87
+ "structure_local_rerank_score_scale": 1.0,
88
+ "structure_logit_margin_mean": -1.0078125,
89
+ "structure_margin_loss": 0.0,
90
+ "structure_margin_loss_weight": 0.0,
91
+ "structure_missing_formula_fraction": 0.0,
92
+ "structure_missing_inchikey_fraction": 0.0,
93
+ "structure_pairwise_logistic_loss": 0.0,
94
+ "structure_pairwise_logistic_loss_weight": 0.0,
95
+ "structure_positive_count_mean": 1.0,
96
+ "structure_post_label_candidate_count": 2126.0,
97
+ "structure_post_precursor_candidate_count": 2128.0,
98
+ "structure_prefilter_candidate_count": 2128.0,
99
+ "structure_retrieval_logit_std": 0.9839304089546204,
100
+ "structure_retrieval_loss": 10.630162239074707,
101
+ "structure_retrieval_scorer": "projected_dot",
102
+ "structure_retrieval_target_mode": "sample",
103
+ "structure_skipped_query_fraction": 0.0,
104
+ "structure_soft_target_exact_mix": 0.4,
105
+ "structure_soft_target_loss": 7.523618221282959,
106
+ "structure_soft_target_loss_weight": 0.3,
107
+ "structure_target_backend": "rdkit_morgan",
108
+ "structure_target_equivalent_count_mean": 1.0,
109
+ "structure_target_identity_collision_count": 0.0,
110
+ "structure_target_rank_mean": 314.703125,
111
+ "structure_to_ssl_ratio": 217.7124787233957,
112
+ "structure_top1_top2_gap_mean": 0.0986328125,
113
+ "structure_top1_top2_gap_norm_mean": 0.10155132412910461,
114
+ "structure_top_negative_logit_mean": 10.1875,
115
+ "structure_uncertain_fraction": 0.970703125,
116
+ "structure_uncertain_hit_at_1": 0.0040241447277367115,
117
+ "structure_valid_query_fraction": 1.0,
118
+ "test_expected_shard_count": 2.0,
119
+ "test_full_coverage": 0.0,
120
+ "test_observed_shard_count": 0.0,
121
+ "train_loss": 6.491608619689941,
122
+ "train_loss_ema": 6.519052214060582,
123
+ "train_retrieval_bank_size": 4096.0,
124
+ "val_deep_eval_single_view": 0.0,
125
+ "val_deep_eval_structure_only": 0.0,
126
+ "val_effective_structure_bce_weight": 1.0,
127
+ "val_eval_loader_mode": "stream",
128
+ "val_eval_mode": "deep",
129
+ "val_expected_shard_count": 2.0,
130
+ "val_full_coverage": 1.0,
131
+ "val_hit_at_1": 0.0282135009765625,
132
+ "val_hit_at_20": 0.15270233154296875,
133
+ "val_hit_at_5": 0.0726776123046875,
134
+ "val_loss": 6.646750174145564,
135
+ "val_observed_eval_batches": 256.0,
136
+ "val_observed_shard_count": 2.0,
137
+ "val_ssl_loss": 1.0397093277424574,
138
+ "val_structure_bce_loss": 0.06532714269997086,
139
+ "val_structure_candidate_count": 4119.69140625,
140
+ "val_structure_clear_fraction": 0.0,
141
+ "val_structure_clear_hit_at_1": 0.0,
142
+ "val_structure_close_fraction": 0.0077972412109375,
143
+ "val_structure_close_hit_at_1": 0.014835291571216658,
144
+ "val_structure_confidence_candidate_count": 4096.0,
145
+ "val_structure_confidence_separation": -0.027957430157584895,
146
+ "val_structure_correct_logit_mean": 9.49462890625,
147
+ "val_structure_fingerprint_cosine": 0.4255251376889646,
148
+ "val_structure_logit_margin_mean": -1.221923828125,
149
+ "val_structure_missing_formula_fraction": 0.0,
150
+ "val_structure_missing_inchikey_fraction": 0.0,
151
+ "val_structure_retrieval_logit_std": 2.6622845763340592,
152
+ "val_structure_retrieval_loss": 10.882395941764116,
153
+ "val_structure_target_equivalent_count_mean": 1.0,
154
+ "val_structure_target_identity_collision_count": 23.69140625,
155
+ "val_structure_target_rank_mean": 379.8781280517578,
156
+ "val_structure_top1_top2_gap_mean": 0.11855697631835938,
157
+ "val_structure_top1_top2_gap_norm_mean": 0.05411380449368153,
158
+ "val_structure_top_negative_logit_mean": 10.714599609375,
159
+ "val_structure_uncertain_fraction": 0.9922027587890625,
160
+ "val_structure_uncertain_hit_at_1": 0.027957430157584895,
161
+ "variance_loss": 0.00213623046875,
162
+ "variance_weight": 0.8,
163
+ "weighted_covariance_loss": 0.015018927097320557,
164
+ "weighted_regularizer_to_contrastive_ratio": 0.028996589606316913,
165
+ "weighted_ssl_loss": 0.02968101501464844,
166
+ "weighted_structure_bce_loss": 0.08383000642061234,
167
+ "weighted_structure_retrieval_loss": 6.378097343444824,
168
+ "weighted_variance_loss": 0.001708984375
169
+ }
metadata/perf_summary.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "logged_points": 521,
3
+ "train_loss_min": 3.2278988361358643,
4
+ "train_loss_max": 7.677141189575195,
5
+ "train_loss_last": 6.491608619689941,
6
+ "train_loss_ema_last": 6.519052214060582,
7
+ "weighted_regularizer_ratio_median": 0.02686493425176378,
8
+ "weighted_regularizer_ratio_p90": 0.09558363904110317,
9
+ "samples_per_sec_mean": 1545.3851888056406,
10
+ "samples_per_sec_min": 930.5498563655322,
11
+ "samples_per_sec_max": 2247.460787425693,
12
+ "step_time_mean": 0.33468720350667036,
13
+ "mfu_mean": 0.06923433462788683,
14
+ "mfu_max": 0.10068781126337953,
15
+ "gpu_mem_max_gb": 13.279157161712646
16
+ }
metadata/results.md ADDED
@@ -0,0 +1,171 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # MS/MS Training Run Result
2
+
3
+ - Run name: `final_v3_structure_alignment_v26_r1`
4
+ - Created at: `2026-04-13T21:36:19Z`
5
+ - Dataset root: `/home/ubuntu/data/gems_v1_handoff_labeled_v26`
6
+ - Shard counts: `{'train': 4, 'val': 2, 'test': 2}`
7
+
8
+ ## Primary Metrics
9
+
10
+ - `train_loss`: `6.491608619689941`
11
+ - `val_loss`: `6.646750174145564`
12
+ - `grad_norm`: `1.7659728527069092`
13
+ - `contrastive_loss`: `0.5768923759460449`
14
+ - `structure_bce_loss`: `0.08383000642061234`
15
+ - `val_structure_bce_loss`: `0.06532714269997086`
16
+ - `val_structure_fingerprint_cosine`: `0.4255251376889646`
17
+ - `different_inchikey_mean_cosine`: `0.004679911238535274`
18
+ - `embedding_std`: `0.07216806709766388`
19
+ - `val_hit_at_20`: `0.15270233154296875`
20
+ - `val_structure_candidate_count`: `4119.69140625`
21
+ - `val_structure_top1_top2_gap_mean`: `0.11855697631835938`
22
+ - `val_structure_uncertain_fraction`: `0.9922027587890625`
23
+ - `val_structure_clear_hit_at_1`: `0.0`
24
+ - `val_structure_confidence_separation`: `-0.027957430157584895`
25
+ - `samples_per_sec`: `1399.0203442084614`
26
+
27
+ ## Failure Mode Assessment
28
+
29
+ - `dataset_shards_accessible`: pass. Observed shard counts {'train': 4, 'val': 2, 'test': 2}; the canonical Day 1 loader contract requires all splits to be readable.
30
+ - `dataloader_streaming_stable`: pass. Training batches continued to arrive with positive throughput and no iterator stall.
31
+ - `non_finite_loss_or_tensor_failure`: pass. Technical_Plan requires zero NaN or Inf tensors and finite train/val loss.
32
+ - `validation_observability_gap`: pass. Engineering.md expects slope-aware monitoring; missing validation or retrieval metrics make trend judgment weaker.
33
+ - `gpu_memory_instability`: pass. Technical_Plan Day 1 success requires stable GPU memory through the bounded run.
34
+ - `checkpoint_save_reload_failure`: pass. Checkpoint save and reload is an explicit Day 1 success criterion.
35
+ - `embedding_collapse`: pass. Engineering.md expects embedding variance monitoring and non-collapsed representations.
36
+ - `regularizer_dominance`: pass. The current instability mode is regularization overpowering the SSL objective and producing jagged total loss.
37
+ - `structure_supervision_gap`: pass. Once the bridge phase begins, Technical_Plan requires retrieval-oriented validation rather than only augmented contrastive loss.
38
+ - `confidence_surface_gap`: pass. Structure-first qualification now requires explicit confidence diagnostics such as top-1/top-2 separation and uncertainty buckets, not just broader retrieval metrics.
39
+ - `chemistry_target_backend_gap`: pass. The bridge phase should prefer chemistry-grounded fingerprint targets; hashed fallback targets weaken the retrieval signal.
40
+ - `bridge_fit_without_ranking`: pass. Bridge runs must improve retrieval after fingerprint fit saturates; fitting BCE while val Hit@20 stays at candidate-bank baseline means ranking is not generalizing.
41
+ - `metadata_shortcutting`: watch. Engineering.md flags acquisition-artefact shortcut risk; adduct and instrument embeddings remain deliberately down-scaled.
42
+
43
+ ## Pressure Summary
44
+
45
+ - all metrics overall: `↗ improving (+21.89% weighted)`
46
+ - all metrics weighted improvement: `+21.89%`
47
+ - all metrics average improvement: `+34.52%`
48
+ - all metrics weighted pressure score: `0.284`
49
+ - all metrics counts: `up=21` `none=7` `down=11` `up_fraction=+53.85%`
50
+ - primary metrics overall: `↗ improving (-4.45% weighted)`
51
+ - primary metrics weighted improvement: `-4.45%`
52
+
53
+ ## Metric Pressure
54
+
55
+ - `train_loss`: `goal ↓ | ↗ +10.57% improving` (early `6.340343519442343`, mid `5.795929590401622`, late `5.670367905463295`, last `6.491608619689941`)
56
+ - `train_loss_ema`: `goal ↓ | ↗ +10.46% improving` (early `6.265581717362126`, mid `5.8371378655061745`, late `5.610470838300897`, last `6.519052214060582`)
57
+ - `contrastive_loss`: `goal ↓ | → -1.64% flat` (early `1.409792113166324`, mid `1.3598130204084982`, late `1.43291388983014`, last `0.5768923759460449`)
58
+ - `val_loss`: `goal ↓ | ↘ -55.82% worsening` (early `1.2876316868363775`, mid `2.20697861268016`, late `2.006446816829876`, last `6.646750174145564`)
59
+ - `val_ssl_loss`: `goal ↓ | → +2.32% flat` (early `1.0849772497183745`, mid `1.0747967428483425`, late `1.0598136518327963`, last `1.0397093277424574`)
60
+ - `structure_bce_loss`: `goal ↓ | ↗ +39.75% improving` (early `0.12293951650489272`, mid `0.07652380578004556`, late `0.07407209518012302`, last `0.08383000642061234`)
61
+ - `val_structure_bce_loss`: `goal ↓ | ↗ +12.20% improving` (early `0.0771060927848642`, mid `0.07029345531468391`, late `0.06769936999660872`, last `0.06532714269997086`)
62
+ - `structure_retrieval_loss`: `goal ↓ | ↗ +10.09% improving` (early `10.237479978903181`, mid `9.41523380775672`, late `9.204101929719421`, last `10.630162239074707`)
63
+ - `val_structure_retrieval_loss`: `goal ↓ | → +4.86% flat` (early `11.562767315655947`, mid `12.150514367967844`, late `11.000247791409492`, last `10.882395941764116`)
64
+ - `structure_fingerprint_cosine`: `goal ↑ | ↗ +17.59% improving` (early `0.3965519580968543`, mid `0.4464927057310336`, late `0.46629545006943846`, last `0.46021854877471924`)
65
+ - `val_structure_fingerprint_cosine`: `goal ↑ | → +0.17% flat` (early `0.4147872725734487`, mid `0.4204998221830465`, late `0.4154996814322658`, last `0.4255251376889646`)
66
+ - `structure_target_rank_mean`: `goal ↓ | ↗ +58.80% improving` (early `351.84337743858384`, mid `176.16389315390174`, late `144.9534280711207`, last `314.703125`)
67
+ - `val_structure_target_rank_mean`: `goal ↓ | ↗ +45.45% improving` (early `760.7540435791016`, mid `683.3385162353516`, late `415.0084457397461`, last `379.8781280517578`)
68
+ - `structure_correct_logit_mean`: `goal ↑ | ↗ +47.31% improving` (early `6.134031791907514`, mid `9.474891618497109`, late `9.035739942528735`, last `9.1875`)
69
+ - `val_structure_correct_logit_mean`: `goal ↑ | ↗ +16.44% improving` (early `8.0311279296875`, mid `8.441162109375`, late `9.3515625`, last `9.49462890625`)
70
+ - `structure_top_negative_logit_mean`: `goal ↓ | ↘ -39.72% worsening` (early `7.262102601156069`, mid `10.490606936416185`, late `10.14691091954023`, last `10.1875`)
71
+ - `val_structure_top_negative_logit_mean`: `goal ↓ | ↘ -16.08% worsening` (early `9.2255859375`, mid `10.51171875`, late `10.708740234375`, last `10.714599609375`)
72
+ - `structure_logit_margin_mean`: `goal ↑ | → +1.23% flat` (early `-1.1263999277456647`, mid `-1.0154725885115607`, late `-1.1125740840517242`, last `-1.0078125`)
73
+ - `val_structure_logit_margin_mean`: `goal ↑ | ↘ -13.67% worsening` (early `-1.194091796875`, mid `-2.0707931518554688`, late `-1.3573150634765625`, last `-1.221923828125`)
74
+ - `structure_retrieval_logit_std`: `goal ↑ | ↗ +112.06% improving` (early `1.0455416592559375`, mid `2.0036371215919537`, late `2.2171744848119803`, last `0.9839304089546204`)
75
+ - `val_structure_retrieval_logit_std`: `goal ↑ | ↗ +102.33% improving` (early `1.2326001720502973`, mid `2.022283688886091`, late `2.49387572822161`, last `2.6622845763340592`)
76
+ - `structure_top1_top2_gap_mean`: `goal ↑ | ↗ +59.22% improving` (early `0.09793567106213873`, mid `0.12284365968208093`, late `0.1559343009159483`, last `0.0986328125`)
77
+ - `val_structure_top1_top2_gap_mean`: `goal ↑ | ↗ +54.91% improving` (early `0.07839679718017578`, mid `0.1310901641845703`, late `0.12144279479980469`, last `0.11855697631835938`)
78
+ - `structure_top1_top2_gap_norm_mean`: `goal ↑ | ↘ -35.65% worsening` (early `0.12215662799301864`, mid `0.07032911504236604`, late `0.07860379065547524`, last `0.10155132412910461`)
79
+ - `val_structure_top1_top2_gap_norm_mean`: `goal ↑ | ↘ -16.83% worsening` (early `0.06950851557485294`, mid `0.07577372609375743`, late `0.0578127475164365`, last `0.05411380449368153`)
80
+ - `structure_uncertain_fraction`: `goal ↓ | ↘ -6.11% worsening` (early `0.924663565751445`, mid `0.984747561416185`, late `0.9811422413793104`, last `0.970703125`)
81
+ - `val_structure_uncertain_fraction`: `goal ↓ | → -0.03% flat` (early `0.9941177368164062`, mid `0.9762382507324219`, late `0.9944419860839844`, last `0.9922027587890625`)
82
+ - `structure_confidence_separation`: `goal ↑ | ↘ -531.56% worsening` (early `-0.007918493679557749`, mid `-0.032136323077151974`, late `-0.05001017991881603`, last `-0.0040241447277367115`)
83
+ - `val_structure_confidence_separation`: `goal ↑ | ↘ -204.85% worsening` (early `-0.00795668972114072`, mid `-0.01309866098972634`, late `-0.024256365849851136`, last `-0.027957430157584895`)
84
+ - `structure_hit_at_1`: `goal ↑ | ↗ +549.68% improving` (early `0.007880238439306358`, mid `0.032706376445086706`, late `0.051196569683908046`, last `0.00390625`)
85
+ - `structure_hit_at_5`: `goal ↑ | ↗ +402.47% improving` (early `0.03360955563583815`, mid `0.11673591040462428`, late `0.16887796336206898`, last `0.029296875`)
86
+ - `structure_hit_at_20`: `goal ↑ | ↗ +239.46% improving` (early `0.10721865968208093`, mid `0.28105807442196534`, late `0.3639659662356322`, last `0.1171875`)
87
+ - `val_hit_at_1`: `goal ↑ | → +207.79% flat` (early `0.0079345703125`, mid `0.01372528076171875`, late `0.02442169189453125`, last `0.0282135009765625`)
88
+ - `val_hit_at_5`: `goal ↑ | ↗ +134.85% improving` (early `0.0288543701171875`, mid `0.0425567626953125`, late `0.0677642822265625`, last `0.0726776123046875`)
89
+ - `val_hit_at_20`: `goal ↑ | ↗ +76.67% improving` (early `0.08158111572265625`, mid `0.094146728515625`, late `0.14413070678710938`, last `0.15270233154296875`)
90
+ - `weighted_regularizer_to_contrastive_ratio`: `goal ↓ | ↗ +54.61% improving` (early `0.0667222202939108`, mid `0.0329437448262181`, late `0.03028750823471251`, last `0.028996589606316913`)
91
+ - `samples_per_sec`: `goal ↑ | ↘ -2.88% worsening` (early `1575.5664566260843`, mid `1531.3735887168505`, late `1530.1496274666802`, last `1399.0203442084614`)
92
+ - `gpu_max_memory_gb`: `goal ≈ | ↗ +100.00% improving` (early `13.252925169950275`, mid `13.279059082097401`, late `13.279157161712646`, last `13.279157161712646`)
93
+ - `grad_norm`: `goal ≈ | ↘ -100.00% worsening` (early `4.638080189338309`, mid `4.979553528603791`, late `6.072961742165445`, last `1.7659728527069092`)
94
+
95
+ ## Primary Metric Pressure
96
+
97
+ - `train_loss`: `goal ↓ | ↗ +10.57% improving` (early `6.340343519442343`, mid `5.795929590401622`, late `5.670367905463295`, last `6.491608619689941`)
98
+ - `val_loss`: `goal ↓ | ↘ -55.82% worsening` (early `1.2876316868363775`, mid `2.20697861268016`, late `2.006446816829876`, last `6.646750174145564`)
99
+ - `grad_norm`: `goal ≈ | ↘ -100.00% worsening` (early `4.638080189338309`, mid `4.979553528603791`, late `6.072961742165445`, last `1.7659728527069092`)
100
+ - `contrastive_loss`: `goal ↓ | → -1.64% flat` (early `1.409792113166324`, mid `1.3598130204084982`, late `1.43291388983014`, last `0.5768923759460449`)
101
+ - `structure_bce_loss`: `goal ↓ | ↗ +39.75% improving` (early `0.12293951650489272`, mid `0.07652380578004556`, late `0.07407209518012302`, last `0.08383000642061234`)
102
+ - `val_structure_bce_loss`: `goal ↓ | ↗ +12.20% improving` (early `0.0771060927848642`, mid `0.07029345531468391`, late `0.06769936999660872`, last `0.06532714269997086`)
103
+ - `val_structure_fingerprint_cosine`: `goal ↑ | → +0.17% flat` (early `0.4147872725734487`, mid `0.4204998221830465`, late `0.4154996814322658`, last `0.4255251376889646`)
104
+ - `val_hit_at_20`: `goal ↑ | ↗ +76.67% improving` (early `0.08158111572265625`, mid `0.094146728515625`, late `0.14413070678710938`, last `0.15270233154296875`)
105
+ - `val_structure_top1_top2_gap_mean`: `goal ↑ | ↗ +54.91% improving` (early `0.07839679718017578`, mid `0.1310901641845703`, late `0.12144279479980469`, last `0.11855697631835938`)
106
+ - `val_structure_uncertain_fraction`: `goal ↓ | → -0.03% flat` (early `0.9941177368164062`, mid `0.9762382507324219`, late `0.9944419860839844`, last `0.9922027587890625`)
107
+ - `val_structure_confidence_separation`: `goal ↑ | ↘ -204.85% worsening` (early `-0.00795668972114072`, mid `-0.01309866098972634`, late `-0.024256365849851136`, last `-0.027957430157584895`)
108
+ - `samples_per_sec`: `goal ↑ | ↘ -2.88% worsening` (early `1575.5664566260843`, mid `1531.3735887168505`, late `1530.1496274666802`, last `1399.0203442084614`)
109
+
110
+ ## Debug Metrics
111
+
112
+ - `train_loss`: `6.491608619689941`
113
+ - `train_loss_ema`: `6.519052214060582`
114
+ - `val_loss`: `6.646750174145564`
115
+ - `contrastive_loss`: `0.5768923759460449`
116
+ - `val_ssl_loss`: `1.0397093277424574`
117
+ - `variance_loss`: `0.00213623046875`
118
+ - `covariance_loss`: `10.012618064880371`
119
+ - `structure_bce_loss`: `0.08383000642061234`
120
+ - `structure_retrieval_loss`: `10.630162239074707`
121
+ - `val_structure_retrieval_loss`: `10.882395941764116`
122
+ - `structure_fingerprint_cosine`: `0.46021854877471924`
123
+ - `structure_hit_at_1`: `0.00390625`
124
+ - `structure_hit_at_5`: `0.029296875`
125
+ - `structure_hit_at_20`: `0.1171875`
126
+ - `structure_target_rank_mean`: `314.703125`
127
+ - `structure_logit_margin_mean`: `-1.0078125`
128
+ - `structure_retrieval_logit_std`: `0.9839304089546204`
129
+ - `val_structure_retrieval_logit_std`: `2.6622845763340592`
130
+ - `structure_correct_logit_mean`: `9.1875`
131
+ - `structure_top_negative_logit_mean`: `10.1875`
132
+ - `val_structure_correct_logit_mean`: `9.49462890625`
133
+ - `val_structure_top_negative_logit_mean`: `10.714599609375`
134
+ - `structure_top1_top2_gap_mean`: `0.0986328125`
135
+ - `structure_top1_top2_gap_norm_mean`: `0.10155132412910461`
136
+ - `structure_confidence_candidate_count`: `2126.0`
137
+ - `structure_close_fraction`: `0.029296875`
138
+ - `structure_uncertain_fraction`: `0.970703125`
139
+ - `structure_uncertain_hit_at_1`: `0.0040241447277367115`
140
+ - `structure_confidence_separation`: `-0.0040241447277367115`
141
+ - `val_structure_top1_top2_gap_mean`: `0.11855697631835938`
142
+ - `val_structure_top1_top2_gap_norm_mean`: `0.05411380449368153`
143
+ - `val_structure_confidence_candidate_count`: `4096.0`
144
+ - `val_structure_close_fraction`: `0.0077972412109375`
145
+ - `val_structure_uncertain_fraction`: `0.9922027587890625`
146
+ - `val_structure_close_hit_at_1`: `0.014835291571216658`
147
+ - `val_structure_uncertain_hit_at_1`: `0.027957430157584895`
148
+ - `val_structure_confidence_separation`: `-0.027957430157584895`
149
+ - `grad_norm`: `1.7659728527069092`
150
+ - `weighted_regularizer_to_contrastive_ratio`: `0.028996589606316913`
151
+ - `samples_per_sec`: `1399.0203442084614`
152
+ - `samples_per_sec_60s`: `1399.0203442084614`
153
+ - `samples_per_sec_300s`: `1547.033197547462`
154
+ - `samples_per_sec_900s`: `1512.7885250414204`
155
+ - `samples_per_sec_run_mean`: `1533.0231795943853`
156
+ - `gpu_max_memory_gb`: `13.279157161712646`
157
+ - `gpu_util_mean_pct`: `36.0`
158
+ - `gpu_util_min_pct`: `18.0`
159
+ - `gpu_util_max_pct`: `54.0`
160
+ - `gpu_active_fraction`: `1.0`
161
+ - `gpu_mem_used_mean_gb`: `14.1943359375`
162
+ - `gpu_mem_util_mean_pct`: `14.848754175733243`
163
+ - `gpu_snapshot_available`: `1.0`
164
+ - `step_time_sec`: `0.3659703749981418`
165
+ - `step_time_cv_300s`: `0.06743440395418321`
166
+ - `step_time_cv_900s`: `0.09218780075615579`
167
+ - `dataloader_wait_pct_60s`: `0.0961140629935623`
168
+ - `dataloader_wait_pct_300s`: `0.0382592069912683`
169
+ - `dataloader_wait_pct_900s`: `0.066640439237091`
170
+ - `dataloader_wait_pct_run`: `0.06717802801040498`
171
+ - `mfu_estimate`: `0.06267708747552383`
metadata/training_surface.json ADDED
@@ -0,0 +1,68 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "created_at": "2026-04-13T20:20:55Z",
3
+ "run_name": "final_v3_structure_alignment_v26_r1",
4
+ "wandb_project": "Nexa_Mol",
5
+ "wandb_run_name": "final_v3_structure_alignment_v26_r1",
6
+ "wandb_group": "Structure-Alignment",
7
+ "dataset_root": "/home/ubuntu/data/gems_v1_handoff_labeled_v26",
8
+ "selected_shards_json": "/home/ubuntu/data/gems_v1_handoff_labeled_v26/selected_shards.json",
9
+ "selected_shards_exists": true,
10
+ "phase_ready_marker": "",
11
+ "phase_ready_exists": false,
12
+ "phase_receipt_path": "",
13
+ "phase_receipt_exists": false,
14
+ "train_shard_surface_mode": "train_only",
15
+ "refresh_selected_shards_on_iterator_reset": false,
16
+ "parent_policy": "local_only",
17
+ "labeled_eval_root": "",
18
+ "query_dataset_root": "",
19
+ "bank_dataset_root": "",
20
+ "primary_checkpoint_run_name": "",
21
+ "fallback_checkpoint_run_name": "",
22
+ "local_parent_checkpoint": "/home/ubuntu/artifacts/nexa_mol/v1_full_pretrain/phase1_blackwell_v3/Final_V3.pt",
23
+ "current_checkpoint_path": "",
24
+ "current_checkpoint_exists": true,
25
+ "disable_auto_resume": true,
26
+ "resume_model_only": true,
27
+ "objective": "hybrid",
28
+ "ssl_weight": 0.05,
29
+ "structure_weight": 1.0,
30
+ "structure_post_saturation_weight": 0.0,
31
+ "structure_threshold": 0.0,
32
+ "structure_lock_patience": 0,
33
+ "train_retrieval_watch_only": false,
34
+ "freeze_retrieval_head": false,
35
+ "retrieval_weight": 0.6,
36
+ "batch_size": 512,
37
+ "eval_batch_size": 512,
38
+ "num_workers": 12,
39
+ "grad_accum_steps": 1,
40
+ "loader_prefetch_factor": 4,
41
+ "log_every": 20,
42
+ "sync_every": 1000,
43
+ "amp_mode": "",
44
+ "additional_steps": 10000,
45
+ "max_steps": 91001,
46
+ "eval_batches": 24,
47
+ "eval_batches_per_shard": 0,
48
+ "eval_start_after_steps": 0,
49
+ "deep_eval_every": 2000,
50
+ "deep_eval_start_after_steps": 2000,
51
+ "deep_eval_batches": 256,
52
+ "deep_eval_single_view": false,
53
+ "deep_eval_structure_only": false,
54
+ "active_train_shards": 0,
55
+ "prefetch_train_shards": 0,
56
+ "val_shards": 0,
57
+ "test_shards": 0,
58
+ "train_offset": 0,
59
+ "val_offset": 0,
60
+ "test_offset": 0,
61
+ "selected_shards_summary": {
62
+ "train_count": 4,
63
+ "val_count": 2,
64
+ "test_count": 2,
65
+ "prefetch_count": 0,
66
+ "subset_strategy": "v26_native_confidence_policy_subset"
67
+ }
68
+ }
model_index.json ADDED
@@ -0,0 +1,48 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "model-index": [
3
+ {
4
+ "name": "NexaMass-V3-Struct",
5
+ "results": [
6
+ {
7
+ "dataset": {
8
+ "name": "GeMS-derived labeled validation surface",
9
+ "type": "roman-bushuiev/GeMS"
10
+ },
11
+ "metrics": [
12
+ {
13
+ "name": "Validation structure BCE",
14
+ "type": "binary_cross_entropy",
15
+ "value": 0.0653
16
+ },
17
+ {
18
+ "name": "Validation fingerprint cosine",
19
+ "type": "cosine_similarity",
20
+ "value": 0.4255
21
+ }
22
+ ],
23
+ "task": {
24
+ "name": "MS/MS structure-aware representation learning",
25
+ "type": "feature-extraction"
26
+ }
27
+ },
28
+ {
29
+ "dataset": {
30
+ "name": "MassSpecGym test dataloader",
31
+ "type": "massspecgym"
32
+ },
33
+ "metrics": [
34
+ {
35
+ "name": "Hit@20",
36
+ "type": "hit@20",
37
+ "value": 0.3505
38
+ }
39
+ ],
40
+ "task": {
41
+ "name": "MassSpecGym adapter sanity check",
42
+ "type": "retrieval"
43
+ }
44
+ }
45
+ ]
46
+ }
47
+ ]
48
+ }
runtime/README.md ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # NexaMass Runtime
2
+
3
+ `nexamass_encoder.py` is a minimal PyTorch loader for the model-state files in this repository.
4
+
5
+ Recommended checkpoint for inference:
6
+
7
+ ```python
8
+ from runtime.nexamass_encoder import ModelConfig, load_nexamass_model_state
9
+
10
+ model = load_nexamass_model_state("weights/NexaMass-V3-Struct-model_state.pt", ModelConfig()) # PyTorch fallback
11
+ # Safetensors weights are also provided for safer tensor-only loading: weights/NexaMass-V3-Struct-model_state.safetensors
12
+ ```
13
+
14
+ Use `weights/NexaMass-V3-Struct-model_state.safetensors` for the RDKit/Morgan-aligned model and `weights/Final_V3-model_state.safetensors` for the self-supervised foundation parent. PyTorch model-state `.pt` fallbacks are also included. Optimizer-bearing full training checkpoints remain in object storage and are intentionally excluded from this public HF payload.
runtime/nexamass_encoder.py ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from __future__ import annotations
2
+
3
+ from dataclasses import dataclass
4
+
5
+ import torch
6
+ import torch.nn as nn
7
+ import torch.nn.functional as F
8
+
9
+
10
+ @dataclass(frozen=True)
11
+ class ModelConfig:
12
+ max_peaks: int = 256
13
+ mz_max: float = 2000.0
14
+ collision_max: float = 200.0
15
+ model_dim: int = 384
16
+ layers: int = 6
17
+ heads: int = 8
18
+ dropout: float = 0.1
19
+ projection_dim: int = 192
20
+ fingerprint_dim: int = 2048
21
+ target_projection_dim: int = 256
22
+ retrieval_mlp_hidden_dim: int = 512
23
+ metadata_scale: float = 0.02
24
+
25
+
26
+ class NexaMassSpectralEncoder(nn.Module):
27
+ """Encoder-only MS/MS transformer used by NexaMass-V3-Struct.
28
+
29
+ Expected batch keys:
30
+ - mzs, ints, mz_to_precursor, peak_rank: float tensors [batch, max_peaks]
31
+ - precursor_mz, charge, collision_energy, peak_count: float tensors [batch]
32
+ - adduct_id, instrument_id: long tensors [batch]
33
+ - mask: bool tensor [batch, max_peaks], True for valid peaks
34
+ """
35
+
36
+ def __init__(self, cfg: ModelConfig) -> None:
37
+ super().__init__()
38
+ self.cfg = cfg
39
+ self.adduct_embedding = nn.Embedding(64, cfg.model_dim)
40
+ self.instrument_embedding = nn.Embedding(64, cfg.model_dim)
41
+ self.input_projection = nn.Linear(8, cfg.model_dim)
42
+ encoder_layer = nn.TransformerEncoderLayer(
43
+ d_model=cfg.model_dim,
44
+ nhead=cfg.heads,
45
+ dim_feedforward=cfg.model_dim * 4,
46
+ dropout=cfg.dropout,
47
+ activation="gelu",
48
+ batch_first=True,
49
+ norm_first=True,
50
+ )
51
+ try:
52
+ self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=cfg.layers, enable_nested_tensor=False)
53
+ except TypeError:
54
+ self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=cfg.layers)
55
+ self.final_norm = nn.LayerNorm(cfg.model_dim)
56
+ self.projection = nn.Sequential(
57
+ nn.Linear(cfg.model_dim, cfg.model_dim),
58
+ nn.GELU(),
59
+ nn.Dropout(cfg.dropout),
60
+ nn.Linear(cfg.model_dim, cfg.projection_dim),
61
+ )
62
+ self.structure_head = nn.Sequential(
63
+ nn.Linear(cfg.model_dim, cfg.model_dim),
64
+ nn.GELU(),
65
+ nn.Dropout(cfg.dropout),
66
+ nn.Linear(cfg.model_dim, cfg.fingerprint_dim),
67
+ )
68
+ self.structure_query = nn.Sequential(
69
+ nn.Linear(cfg.model_dim, cfg.model_dim),
70
+ nn.GELU(),
71
+ nn.Dropout(cfg.dropout),
72
+ nn.Linear(cfg.model_dim, cfg.target_projection_dim),
73
+ )
74
+ self.target_projection = nn.Sequential(
75
+ nn.Linear(cfg.fingerprint_dim, cfg.model_dim),
76
+ nn.GELU(),
77
+ nn.Dropout(cfg.dropout),
78
+ nn.Linear(cfg.model_dim, cfg.target_projection_dim),
79
+ )
80
+ self.retrieval_bilinear = nn.Linear(cfg.target_projection_dim, cfg.target_projection_dim, bias=False)
81
+ self.retrieval_pair_mlp = nn.Sequential(
82
+ nn.Linear(cfg.target_projection_dim * 4, cfg.retrieval_mlp_hidden_dim),
83
+ nn.GELU(),
84
+ nn.Dropout(cfg.dropout),
85
+ nn.Linear(cfg.retrieval_mlp_hidden_dim, 1),
86
+ )
87
+ self.local_rerank_mlp = nn.Sequential(
88
+ nn.Linear(cfg.target_projection_dim * 4 + 1, cfg.retrieval_mlp_hidden_dim),
89
+ nn.GELU(),
90
+ nn.Dropout(cfg.dropout),
91
+ nn.Linear(cfg.retrieval_mlp_hidden_dim, 1),
92
+ )
93
+
94
+ def encode(self, batch: dict[str, torch.Tensor]) -> torch.Tensor:
95
+ features = torch.stack(
96
+ [
97
+ batch["mzs"],
98
+ batch["ints"],
99
+ batch["mz_to_precursor"],
100
+ batch["peak_rank"],
101
+ batch["precursor_mz"].unsqueeze(-1).expand_as(batch["mzs"]),
102
+ batch["charge"].unsqueeze(-1).expand_as(batch["mzs"]),
103
+ batch["collision_energy"].unsqueeze(-1).expand_as(batch["mzs"]),
104
+ batch["peak_count"].unsqueeze(-1).expand_as(batch["mzs"]),
105
+ ],
106
+ dim=-1,
107
+ )
108
+ hidden = self.input_projection(features)
109
+ hidden = hidden + self.adduct_embedding(batch["adduct_id"])[:, None, :] * self.cfg.metadata_scale
110
+ hidden = hidden + self.instrument_embedding(batch["instrument_id"])[:, None, :] * self.cfg.metadata_scale
111
+ encoded = self.encoder(hidden, src_key_padding_mask=~batch["mask"])
112
+ encoded = self.final_norm(encoded)
113
+ mask = batch["mask"].unsqueeze(-1)
114
+ return (encoded * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1)
115
+
116
+ def forward_with_heads(
117
+ self, batch: dict[str, torch.Tensor]
118
+ ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
119
+ pooled = self.encode(batch)
120
+ raw_projected = self.projection(pooled)
121
+ structure_logits = self.structure_head(pooled)
122
+ structure_query_raw = self.structure_query(pooled)
123
+ return F.normalize(raw_projected, dim=-1), raw_projected, structure_logits, structure_query_raw
124
+
125
+ def project_structure_targets(self, targets: torch.Tensor) -> torch.Tensor:
126
+ return F.normalize(self.target_projection(targets), dim=-1)
127
+
128
+
129
+ def load_nexamass_model_state(
130
+ checkpoint_path: str,
131
+ cfg: ModelConfig | None = None,
132
+ map_location: str | torch.device = "cpu",
133
+ ) -> NexaMassSpectralEncoder:
134
+ payload = torch.load(checkpoint_path, map_location=map_location)
135
+ cfg = cfg or ModelConfig()
136
+ model = NexaMassSpectralEncoder(cfg)
137
+ model.load_state_dict(payload["model_state"], strict=True)
138
+ model.eval()
139
+ return model
weights/Final_V3-model_state.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c16f3bb2ab7c2612b6b84a3359abeaf0a1cb1c055399beac21b39521424328b5
3
+ size 56473018
weights/Final_V3-model_state.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:870b30c7bb6a340a460306882fe631d76282f233baf3920ca3a45c363eaf0ee6
3
+ size 56437128
weights/NexaMass-V3-Struct-model_state.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bde396a7448b6972bcb784a465812b4d5b819f8d0c76cf2fba7066db7be9da1f
3
+ size 56474172
weights/NexaMass-V3-Struct-model_state.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e006a6fa4ba6b8f127c7eabc9eaf6b7b64675e0136dee3ca1655bb3a4a3eadd
3
+ size 56437136