Promote v9 preview checkpoint

ESS-AIST-81M.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3b7eb74bacea98e7122e723c164dfd912dd1ccb6902605972f905f95c337dfa2
-size 323643112
+oid sha256:415d6f5ac8299fd17265a6d1ae5ccafeed26729b212231778188dbadceaf6fba
+size 323643096

README.md

@@ -9,7 +9,6 @@ tags:
 - retrieval
 - image-text-audio
 - feature-extraction
-- gguf
 library_name: pytorch
 pipeline_tag: feature-extraction
 datasets:
@@ -20,12 +19,27 @@ datasets:
 
 `ESS-AIST-81M Preview` is the current Cortext trial checkpoint from the ESS line.
 
-- release checkpoint: `ess_aist_full_v7_librispeech360_l4i/checkpoint_epoch_11.pt`
+- release checkpoint: `ess_aist_full_v9_subjectfix_l4k/best_model.pt`
+- exported checkpoint epoch: `3`
 - text encoder: `MongoDB/mdbr-leaf-ir`
 - image encoder: `mobilenetv4_conv_medium.e180_r384_in12k`
 - audio encoder: native `mn20_as` EfficientAT LoRA audio backbone
 
-This is a preview artifact. It restores real speech retrieval and keeps the ESS semantic / subject / event slice layout, but subject continuity is still the weakest domain and remains under active work.
+This preview is the current bridge artifact for Cortext. It keeps the ESS
+`semantic / subject / event` slice layout, but the `v9` dataset repair moved
+the `subject` slice much closer to the entity signal Cortext actually needs.
+
+GGUF quantizations for this exact release live in:
+
+- `augmem/ESS-AIST-81M-preview-GGUF`
+
+Tradeoff:
+
+- held-out subject/entity separation is much stronger than the earlier `v7` preview
+- speech and SALT retrieval are weaker than the earlier `v7` retrieval-max point
+
+For Cortext, this is still the better preview because the entity-side signal is
+materially stronger.
 
 ## Embedding Layout
 
@@ -44,7 +58,8 @@ Recommended normalized runtime views:
 
 ## Exact Release Metrics
 
-All numbers below are from the exact published checkpoint `checkpoint_epoch_11.pt`.
+All numbers below are from the exact published checkpoint state exported from
+`ess_aist_full_v9_subjectfix_l4k/best_model.pt` at checkpoint epoch `3`.
 
 Evaluation scope note:
 
@@ -60,19 +75,19 @@ Source:
 
 Speech holdout:
 
-- `A->T_r1 = 0.4672`
-- `T->A_r1 = 0.4606`
-- `A->T_r5 = 0.7398`
-- `T->A_r5 = 0.7426`
+- `A->T_r1 = 0.3276`
+- `T->A_r1 = 0.3202`
+- `A->T_r5 = 0.6120`
+- `T->A_r5 = 0.6046`
 
 SALT:
 
-- `I->T_r1 = 0.4149`
-- `T->I_r1 = 0.4327`
-- `A->T_r1 = 0.2408`
-- `T->A_r1 = 0.2486`
-- `I->A_r1 = 0.4621`
-- `A->I_r1 = 0.4829`
+- `I->T_r1 = 0.3179`
+- `T->I_r1 = 0.3425`
+- `A->T_r1 = 0.1226`
+- `T->A_r1 = 0.1272`
+- `I->A_r1 = 0.1970`
+- `A->I_r1 = 0.2148`
 
 ### Held-Out ESS Eval
 
@@ -82,22 +97,26 @@ Sources:
 - `event_eval.json`
 - `prefix_eval.json`
 
-Subject:
+Subject / entity surface:
 
-- `subject_key` same/different AUC: `0.5067`
-- `subject_key` same-topic-different-subject rejection AUC: `0.5067`
+- `subject_key` same/different AUC: `0.9881`
+- `subject_key` same-topic-different-subject rejection AUC: `0.9881`
 
-Event:
+Event / disambiguation surface:
 
-- `event_key` same/different AUC: `0.8241`
-- `event_key` same-subject-different-event rejection AUC: `0.5535`
-- `event_key` topic-shift rejection AUC: `0.9770`
+- `subject_key` event same/different AUC: `0.8855`
+- `event_key` event same/different AUC: `0.8193`
+- `subject_key` same-subject-different-event rejection AUC: `0.7381`
+- `event_key` same-subject-different-event rejection AUC: `0.6807`
+- `subject_key` topic-shift rejection AUC: `0.9513`
+- `event_key` topic-shift rejection AUC: `0.8969`
 
 Interpretation:
 
-- speech recovery is real
-- event continuity is usable
-- subject continuity is not yet strong on the current held-out text-anchor eval
+- the repaired `v9` held-out surface is no longer near-random on subject/entity
+- the current `subject` slice is the strongest entity carrier in the model
+- event structure is usable, but still entangled with subject
+- this is the right bridge checkpoint for Cortext, not the final `semantic/entity` architecture
 
 ## Architecture
 
@@ -111,15 +130,14 @@ This preview is a frozen-encoder / trainable-projector stack:
 - text projection params: `8,926,720`
 - total exact loaded params: `80,812,854`
 
-The audio path is not the old dual-audio teacher path. It uses the native audioheavy LoRA EfficientAT backbone that restored speech retrieval for this line.
+The audio path is not the old dual-audio teacher path. It uses the native
+audioheavy LoRA EfficientAT backbone.
 
 ## Files
 
 | File | Purpose |
 |---|---|
 | `ESS-AIST-81M.safetensors` | Full preview release artifact |
-| `ESS-AIST-81M_q8_0.gguf` | Conservative GGUF quantization |
-| `ESS-AIST-81M_q5_1.gguf` | Smaller GGUF quantization |
 | `export_metadata.json` | ESS export contract |
 | `manifest.json` | Release manifest |
 | `parameter_breakdown.json` | Exact parameter accounting |
@@ -131,8 +149,9 @@ The audio path is not the old dual-audio teacher path. It uses the native audioh
 
 ## Caveats
 
-- This is the current preview checkpoint, not the finished ESS subject-memory model.
-- Subject performance is still the weakest domain on the current held-out eval.
-- The current held-out subject eval measures text-anchor separation and under-measures some multimodal subject gains.
+- This is the current preview checkpoint, not the final Cortext model family.
+- The current runtime slices are still named `semantic / subject / event`; the next family will move toward `semantic / entity`.
+- Subject/entity is now strong on the repaired `v9` held-out surface, but event remains entangled and the engine still needs attention over active anchors for weak-reference resolution.
+- Retrieval on `speech holdout` and `SALT` is lower than the earlier `v7` preview.
 - `SALT` and `speech holdout` are useful release gates for this line, but they are no longer fully external benchmarks in the same way they were for the earlier pre-ESS artifacts.
 - Use this for internal Cortext trials, not as the final memory-model release.

ess_ait_86m_spec.yaml

@@ -61,13 +61,13 @@ early_stopping_patience: 8
 log_dir: runs
 benchmark_eval_every_epochs: 1
 
-# Next-run ESS corpus, adding LibriSpeech person-subject rows on top of the
-# v6 subject-media + WIT + speech/wavcaps semantic lane.
-ess_corpus_dir: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360
-ess_train_jsonl: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360/train.jsonl
-ess_val_jsonl: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360/val.jsonl
-ess_train_text_cache: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360/cache/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_train_leaf_ir_text_features.npy
-ess_val_text_cache: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360/cache/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_val_leaf_ir_text_features.npy
+# Next-run ESS corpus, replacing the weak LibriSpeech subject text with
+# identity-prefixed subject rows and book-level hard negatives.
+ess_corpus_dir: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix
+ess_train_jsonl: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix/train.jsonl
+ess_val_jsonl: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix/val.jsonl
+ess_train_text_cache: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix/cache/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix_train_leaf_ir_text_features.npy
+ess_val_text_cache: checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix/cache/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix_val_leaf_ir_text_features.npy
 
 # Multimodal subject-media attachment from the finalized v19 generated bundle.
 ess_subject_media_dataset_dir: checkpoints/ess_ait_86m_20260430T035907Z/ess_subject_media_pilot52_full_v19
@@ -100,8 +100,8 @@ ess_subject_slice: [512, 1024]
 ess_event_slice: [1024, 1536]
 
 # Corpus composition at build time:
-# - train: 219957 semantic / 28026 event / 94446 subject
-# - val:   19954 semantic / 3074 event / 10967 subject
+# - train: 219957 semantic / 121223 event / 188415 subject
+# - val:   19954 semantic / 13891 event / 21012 subject
 #
 # Do not sample raw row frequency. Subject supervision is too small and must be
 # explicitly oversampled to shape the subject block.
@@ -115,7 +115,9 @@ ess_sampling:
   train_dataset_weights:
     speech_chatterbox_150k: 5.0
     wit_entity_subject: 0.25
-    librispeech_subject: 0.01
+    librispeech_subject: 0.02
+    librispeech_subject_lexical: 0.02
+    librispeech_event_lexical: 0.05
   val_family_weights:
     semantic: 0.50
     subject: 0.15
@@ -123,7 +125,9 @@ ess_sampling:
   val_dataset_weights:
     speech_chatterbox_150k: 5.0
     wit_entity_subject: 0.25
-    librispeech_subject: 0.01
+    librispeech_subject: 0.02
+    librispeech_subject_lexical: 0.02
+    librispeech_event_lexical: 0.05
   family_from_active_supervision:
     semantic: semantic
     subject: subject
@@ -136,7 +140,9 @@ ess_sampling:
     - subject rows are intentionally oversampled relative to raw corpus count
     - semantic remains dominant to protect 512d retrieval
     - speech_chatterbox semantic rows are oversampled within semantic because only ~20k rows survive dedupe into v6
-    - librispeech_subject is heavily downweighted within subject so person voice identity helps without flooding the entire subject block
+    - librispeech_subject now carries identity-prefixed text plus book-level hard negatives, so its dataset weight is raised modestly
+    - librispeech_subject_lexical now mirrors that stronger identity text and is also raised modestly within subject
+    - librispeech_event_lexical adds chapter/book lexical event structure without dominating the SALT event anchor
     - event stays high enough to shape prefix_1536 without overwhelming semantic
 
 ess_loss_weights:

event_eval.json

@@ -1,265 +1,265 @@
 {
-  "checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v7_librispeech360_l4i/checkpoint_epoch_11.pt",
+  "checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v9_subjectfix_l4k/best_model.pt",
   "split": "val",
-  "records_path": "/shared/augmem/triembed/checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360/val.jsonl",
+  "records_path": "/shared/augmem/triembed/checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix/val.jsonl",
   "views": {
     "semantic_key": {
       "event_same_different_auc": {
-        "auc": 0.829112461248993,
-        "positive_pairs": 7703,
-        "negative_pairs": 327075,
-        "positive_mean": 0.7533491437897564,
-        "negative_mean": 0.5995114385256625
+        "auc": 0.827451141075551,
+        "positive_pairs": 181625,
+        "negative_pairs": 453539,
+        "positive_mean": 0.8296324522244216,
+        "negative_mean": 0.6291148030260264
       },
       "same_subject_different_event_rejection_auc": {
-        "auc": 0.5802306316888112,
-        "positive_pairs": 7703,
-        "negative_pairs": 118115,
-        "positive_mean": 0.7533491437897564,
-        "negative_mean": 0.7313196869598024
+        "auc": 0.6695206182798321,
+        "positive_pairs": 181625,
+        "negative_pairs": 175759,
+        "positive_mean": 0.8296324522244216,
+        "negative_mean": 0.7356056283701664
       },
       "stale_same_source_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7533491437897564,
+        "positive_mean": 0.8296324522244216,
         "negative_mean": null
       },
       "wrong_active_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7533491437897564,
+        "positive_mean": 0.8296324522244216,
         "negative_mean": null
       },
       "topic_shift_rejection_auc": {
-        "auc": 0.9697933441859231,
-        "positive_pairs": 7703,
-        "negative_pairs": 208960,
-        "positive_mean": 0.7533491437897564,
-        "negative_mean": 0.525006599016673
+        "auc": 0.8994650340280883,
+        "positive_pairs": 181625,
+        "negative_pairs": 340388,
+        "positive_mean": 0.8296324522244216,
+        "negative_mean": 0.5938547173777865
       }
     },
     "subject_key": {
       "event_same_different_auc": {
-        "auc": 0.6676734827239529,
-        "positive_pairs": 7703,
-        "negative_pairs": 327075,
-        "positive_mean": 0.668074605508422,
-        "negative_mean": 0.5629470272292642
+        "auc": 0.8854762132187833,
+        "positive_pairs": 181625,
+        "negative_pairs": 453539,
+        "positive_mean": 0.8070377156889529,
+        "negative_mean": 0.5549017710295172
       },
       "same_subject_different_event_rejection_auc": {
-        "auc": 0.1862661483021773,
-        "positive_pairs": 7703,
-        "negative_pairs": 118115,
-        "positive_mean": 0.668074605508422,
-        "negative_mean": 0.7863739344748515
+        "auc": 0.7381349173591332,
+        "positive_pairs": 181625,
+        "negative_pairs": 175759,
+        "positive_mean": 0.8070377156889529,
+        "negative_mean": 0.6661491519755037
       },
       "stale_same_source_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.668074605508422,
+        "positive_mean": 0.8070377156889529,
         "negative_mean": null
       },
       "wrong_active_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.668074605508422,
+        "positive_mean": 0.8070377156889529,
         "negative_mean": null
       },
       "topic_shift_rejection_auc": {
-        "auc": 0.9397898078829693,
-        "positive_pairs": 7703,
-        "negative_pairs": 208960,
-        "positive_mean": 0.668074605508422,
-        "negative_mean": 0.43665458298485105
+        "auc": 0.9512869887738572,
+        "positive_pairs": 181625,
+        "negative_pairs": 340388,
+        "positive_mean": 0.8070377156889529,
+        "negative_mean": 0.5185315754949175
       }
     },
     "event_key": {
       "event_same_different_auc": {
-        "auc": 0.8240710674869262,
-        "positive_pairs": 7703,
-        "negative_pairs": 327075,
-        "positive_mean": 0.6786398216704803,
-        "negative_mean": 0.4559661049066459
+        "auc": 0.8193492434516296,
+        "positive_pairs": 181625,
+        "negative_pairs": 453539,
+        "positive_mean": 0.8111014214781179,
+        "negative_mean": 0.560698310072904
       },
       "same_subject_different_event_rejection_auc": {
-        "auc": 0.5534970574958717,
-        "positive_pairs": 7703,
-        "negative_pairs": 118115,
-        "positive_mean": 0.6786398216704803,
-        "negative_mean": 0.6430964619714105
+        "auc": 0.6806606788615208,
+        "positive_pairs": 181625,
+        "negative_pairs": 175759,
+        "positive_mean": 0.8111014214781179,
+        "negative_mean": 0.673238928121084
       },
       "stale_same_source_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.6786398216704803,
+        "positive_mean": 0.8111014214781179,
         "negative_mean": null
       },
       "wrong_active_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.6786398216704803,
+        "positive_mean": 0.8111014214781179,
         "negative_mean": null
       },
       "topic_shift_rejection_auc": {
-        "auc": 0.9770134927840807,
-        "positive_pairs": 7703,
-        "negative_pairs": 208960,
-        "positive_mean": 0.6786398216704803,
-        "negative_mean": 0.35019034818428435
+        "auc": 0.8968700248558907,
+        "positive_pairs": 181625,
+        "negative_pairs": 340388,
+        "positive_mean": 0.8111014214781179,
+        "negative_mean": 0.5184466918134688
       }
     },
     "full_key": {
       "event_same_different_auc": {
-        "auc": 0.78280390304866,
-        "positive_pairs": 7703,
-        "negative_pairs": 327075,
-        "positive_mean": 0.7056915993491164,
-        "negative_mean": 0.5454881820918898
+        "auc": 0.8518429254835135,
+        "positive_pairs": 181625,
+        "negative_pairs": 453539,
+        "positive_mean": 0.8179552220574594,
+        "negative_mean": 0.5906047553254696
       },
       "same_subject_different_event_rejection_auc": {
-        "auc": 0.44637572176232837,
-        "positive_pairs": 7703,
-        "negative_pairs": 118115,
-        "positive_mean": 0.7056915993491164,
-        "negative_mean": 0.7219292155613277
+        "auc": 0.6938164439467958,
+        "positive_pairs": 181625,
+        "negative_pairs": 175759,
+        "positive_mean": 0.8179552220574594,
+        "negative_mean": 0.6977020526065576
       },
       "stale_same_source_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7056915993491164,
+        "positive_mean": 0.8179552220574594,
         "negative_mean": null
       },
       "wrong_active_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7056915993491164,
+        "positive_mean": 0.8179552220574594,
         "negative_mean": null
       },
       "topic_shift_rejection_auc": {
-        "auc": 0.9729705121252057,
-        "positive_pairs": 7703,
-        "negative_pairs": 208960,
-        "positive_mean": 0.7056915993491164,
-        "negative_mean": 0.4457545839475431
+        "auc": 0.9288029360785891,
+        "positive_pairs": 181625,
+        "negative_pairs": 340388,
+        "positive_mean": 0.8179552220574594,
+        "negative_mean": 0.553700150467076
       }
     },
     "prefix_512": {
       "event_same_different_auc": {
-        "auc": 0.829112461248993,
-        "positive_pairs": 7703,
-        "negative_pairs": 327075,
-        "positive_mean": 0.7533491437897564,
-        "negative_mean": 0.5995114385256625
+        "auc": 0.827451141075551,
+        "positive_pairs": 181625,
+        "negative_pairs": 453539,
+        "positive_mean": 0.8296324522244216,
+        "negative_mean": 0.6291148030260264
       },
       "same_subject_different_event_rejection_auc": {
-        "auc": 0.5802306316888112,
-        "positive_pairs": 7703,
-        "negative_pairs": 118115,
-        "positive_mean": 0.7533491437897564,
-        "negative_mean": 0.7313196869598024
+        "auc": 0.6695206182798321,
+        "positive_pairs": 181625,
+        "negative_pairs": 175759,
+        "positive_mean": 0.8296324522244216,
+        "negative_mean": 0.7356056283701664
       },
       "stale_same_source_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7533491437897564,
+        "positive_mean": 0.8296324522244216,
         "negative_mean": null
       },
       "wrong_active_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7533491437897564,
+        "positive_mean": 0.8296324522244216,
         "negative_mean": null
       },
       "topic_shift_rejection_auc": {
-        "auc": 0.9697933441859231,
-        "positive_pairs": 7703,
-        "negative_pairs": 208960,
-        "positive_mean": 0.7533491437897564,
-        "negative_mean": 0.525006599016673
+        "auc": 0.8994650340280883,
+        "positive_pairs": 181625,
+        "negative_pairs": 340388,
+        "positive_mean": 0.8296324522244216,
+        "negative_mean": 0.5938547173777865
       }
     },
     "prefix_1024": {
       "event_same_different_auc": {
-        "auc": 0.7453008223026156,
-        "positive_pairs": 7703,
-        "negative_pairs": 327075,
-        "positive_mean": 0.7183707235626442,
-        "negative_mean": 0.5870387001189491
+        "auc": 0.8613090604277244,
+        "positive_pairs": 181625,
+        "negative_pairs": 453539,
+        "positive_mean": 0.8195616576529665,
+        "negative_mean": 0.5971227166902979
       },
       "same_subject_different_event_rejection_auc": {
-        "auc": 0.3570483627258597,
-        "positive_pairs": 7703,
-        "negative_pairs": 118115,
-        "positive_mean": 0.7183707235626442,
-        "negative_mean": 0.7610074661872391
+        "auc": 0.7009197337402358,
+        "positive_pairs": 181625,
+        "negative_pairs": 175759,
+        "positive_mean": 0.8195616576529665,
+        "negative_mean": 0.7044079956623066
       },
       "stale_same_source_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7183707235626442,
+        "positive_mean": 0.8195616576529665,
         "negative_mean": null
       },
       "wrong_active_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7183707235626442,
+        "positive_mean": 0.8195616576529665,
         "negative_mean": null
       },
       "topic_shift_rejection_auc": {
-        "auc": 0.9647611939666115,
-        "positive_pairs": 7703,
-        "negative_pairs": 208960,
-        "positive_mean": 0.7183707235626442,
-        "negative_mean": 0.48870255538236756
+        "auc": 0.934646406192986,
+        "positive_pairs": 181625,
+        "negative_pairs": 340388,
+        "positive_mean": 0.8195616576529665,
+        "negative_mean": 0.5615511370717131
       }
     },
     "prefix_1536": {
       "event_same_different_auc": {
-        "auc": 0.78280390304866,
-        "positive_pairs": 7703,
-        "negative_pairs": 327075,
-        "positive_mean": 0.7056915993491164,
-        "negative_mean": 0.5454881820918898
+        "auc": 0.8518429254835135,
+        "positive_pairs": 181625,
+        "negative_pairs": 453539,
+        "positive_mean": 0.8179552220574594,
+        "negative_mean": 0.5906047553254696
       },
       "same_subject_different_event_rejection_auc": {
-        "auc": 0.44637572176232837,
-        "positive_pairs": 7703,
-        "negative_pairs": 118115,
-        "positive_mean": 0.7056915993491164,
-        "negative_mean": 0.7219292155613277
+        "auc": 0.6938164439467958,
+        "positive_pairs": 181625,
+        "negative_pairs": 175759,
+        "positive_mean": 0.8179552220574594,
+        "negative_mean": 0.6977020526065576
       },
       "stale_same_source_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7056915993491164,
+        "positive_mean": 0.8179552220574594,
         "negative_mean": null
       },
       "wrong_active_rejection_auc": {
         "auc": null,
-        "positive_pairs": 7703,
+        "positive_pairs": 181625,
         "negative_pairs": 0,
-        "positive_mean": 0.7056915993491164,
+        "positive_mean": 0.8179552220574594,
         "negative_mean": null
       },
       "topic_shift_rejection_auc": {
-        "auc": 0.9729705121252057,
-        "positive_pairs": 7703,
-        "negative_pairs": 208960,
-        "positive_mean": 0.7056915993491164,
-        "negative_mean": 0.4457545839475431
+        "auc": 0.9288029360785891,
+        "positive_pairs": 181625,
+        "negative_pairs": 340388,
+        "positive_mean": 0.8179552220574594,
+        "negative_mean": 0.553700150467076
       }
     }
   }

manifest.json

@@ -1,10 +1,6 @@
 {
   "model_id": "ESS-AIST-81M",
-  "trimodal_checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v7_librispeech360_l4i/checkpoint_epoch_11.pt",
+  "trimodal_checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v9_subjectfix_l4k/best_model.pt",
   "audio_checkpoint": "/shared/augmem/triembed/checkpoints/mn20_native_lora_aistmix_audioheavy100k175k175k_continue_from_balanced_20260426T143137Z/latest_model.pt",
-  "safetensors": "/shared/augmem/triembed/dist/ESS-AIST-81M-preview/ESS-AIST-81M.safetensors",
-  "gguf": [
-    "/shared/augmem/triembed/dist/ESS-AIST-81M-preview/ESS-AIST-81M_q8_0.gguf",
-    "/shared/augmem/triembed/dist/ESS-AIST-81M-preview/ESS-AIST-81M_q5_1.gguf"
-  ]
+  "safetensors": "/shared/augmem/triembed/dist/ESS-AIST-81M-preview-hf/ESS-AIST-81M.safetensors"
 }

parameter_breakdown.json

@@ -1,9 +1,9 @@
 {
   "text_encoder": 22861056,
-  "image_encoder": 8434512,
+  "image_encoder": 8502493,
   "audio_encoder": 20639974,
   "image_projection": 9975296,
   "audio_projection": 9975296,
   "text_projection": 8926720,
-  "total_exact_loaded_params": 80812854
+  "total_exact_loaded_params": 80880835
 }

prefix_eval.json

@@ -1,48 +1,48 @@
 {
-  "checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v7_librispeech360_l4i/checkpoint_epoch_11.pt",
+  "checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v9_subjectfix_l4k/best_model.pt",
   "split": "val",
   "views": {
     "semantic_key": {
-      "subject_same_different_auc": 0.4265240470767738,
-      "event_same_different_auc": 0.829112461248993,
-      "same_topic_different_subject_rejection_auc": 0.4265240470767738,
-      "same_subject_different_event_rejection_auc": 0.5802306316888112
+      "subject_same_different_auc": 0.9562558404233674,
+      "event_same_different_auc": 0.827451141075551,
+      "same_topic_different_subject_rejection_auc": 0.9562558404233674,
+      "same_subject_different_event_rejection_auc": 0.6695206182798321
     },
     "subject_key": {
-      "subject_same_different_auc": 0.5066875746523821,
-      "event_same_different_auc": 0.6676734827239529,
-      "same_topic_different_subject_rejection_auc": 0.5066875746523821,
-      "same_subject_different_event_rejection_auc": 0.1862661483021773
+      "subject_same_different_auc": 0.9881162919768391,
+      "event_same_different_auc": 0.8854762132187833,
+      "same_topic_different_subject_rejection_auc": 0.9881162919768391,
+      "same_subject_different_event_rejection_auc": 0.7381349173591332
     },
     "event_key": {
-      "subject_same_different_auc": 0.3832485276953712,
-      "event_same_different_auc": 0.8240710674869262,
-      "same_topic_different_subject_rejection_auc": 0.3832485276953712,
-      "same_subject_different_event_rejection_auc": 0.5534970574958717
+      "subject_same_different_auc": 0.9551271013544805,
+      "event_same_different_auc": 0.8193492434516296,
+      "same_topic_different_subject_rejection_auc": 0.9551271013544805,
+      "same_subject_different_event_rejection_auc": 0.6806606788615208
     },
     "full_key": {
-      "subject_same_different_auc": 0.42067046032727157,
-      "event_same_different_auc": 0.78280390304866,
-      "same_topic_different_subject_rejection_auc": 0.42067046032727157,
-      "same_subject_different_event_rejection_auc": 0.44637572176232837
+      "subject_same_different_auc": 0.9778614751548688,
+      "event_same_different_auc": 0.8518429254835135,
+      "same_topic_different_subject_rejection_auc": 0.9778614751548688,
+      "same_subject_different_event_rejection_auc": 0.6938164439467958
     },
     "prefix_512": {
-      "subject_same_different_auc": 0.4265240470767738,
-      "event_same_different_auc": 0.829112461248993,
-      "same_topic_different_subject_rejection_auc": 0.4265240470767738,
-      "same_subject_different_event_rejection_auc": 0.5802306316888112
+      "subject_same_different_auc": 0.9562558404233674,
+      "event_same_different_auc": 0.827451141075551,
+      "same_topic_different_subject_rejection_auc": 0.9562558404233674,
+      "same_subject_different_event_rejection_auc": 0.6695206182798321
     },
     "prefix_1024": {
-      "subject_same_different_auc": 0.4690923681629257,
-      "event_same_different_auc": 0.7453008223026156,
-      "same_topic_different_subject_rejection_auc": 0.4690923681629257,
-      "same_subject_different_event_rejection_auc": 0.3570483627258597
+      "subject_same_different_auc": 0.9814636892484202,
+      "event_same_different_auc": 0.8613090604277244,
+      "same_topic_different_subject_rejection_auc": 0.9814636892484202,
+      "same_subject_different_event_rejection_auc": 0.7009197337402358
     },
     "prefix_1536": {
-      "subject_same_different_auc": 0.42067046032727157,
-      "event_same_different_auc": 0.78280390304866,
-      "same_topic_different_subject_rejection_auc": 0.42067046032727157,
-      "same_subject_different_event_rejection_auc": 0.44637572176232837
+      "subject_same_different_auc": 0.9778614751548688,
+      "event_same_different_auc": 0.8518429254835135,
+      "same_topic_different_subject_rejection_auc": 0.9778614751548688,
+      "same_subject_different_event_rejection_auc": 0.6938164439467958
     }
   }
 }

retrieval_512_gt1030.json

@@ -1,40 +1,40 @@
 {
   "SALT-512": {
-    "A->I_r1": 0.4828965961933136,
-    "A->I_r10": 0.8761752843856812,
-    "A->I_r5": 0.7863572835922241,
-    "A->T_r1": 0.24084816873073578,
-    "A->T_r10": 0.5153030753135681,
-    "A->T_r5": 0.45209044218063354,
-    "I->A_r1": 0.46209242939949036,
-    "I->A_r10": 0.881176233291626,
-    "I->A_r5": 0.7905581593513489,
-    "I->T_r1": 0.41488298773765564,
-    "I->T_r10": 0.5707141757011414,
-    "I->T_r5": 0.5401080250740051,
-    "T->A_r1": 0.2486497312784195,
-    "T->A_r10": 0.5323064923286438,
-    "T->A_r5": 0.46209242939949036,
-    "T->I_r1": 0.43268653750419617,
-    "T->I_r10": 0.5763152837753296,
-    "T->I_r5": 0.550710141658783
+    "A->I_r1": 0.21484297513961792,
+    "A->I_r10": 0.6691338419914246,
+    "A->I_r5": 0.5125024914741516,
+    "A->T_r1": 0.12262453138828278,
+    "A->T_r10": 0.41028207540512085,
+    "A->T_r5": 0.30946189165115356,
+    "I->A_r1": 0.1970394104719162,
+    "I->A_r10": 0.6443288922309875,
+    "I->A_r5": 0.48849770426750183,
+    "I->T_r1": 0.3178635835647583,
+    "I->T_r10": 0.5503100752830505,
+    "I->T_r5": 0.4920984208583832,
+    "T->A_r1": 0.12722544372081757,
+    "T->A_r10": 0.41768354177474976,
+    "T->A_r5": 0.31926384568214417,
+    "T->I_r1": 0.3424685001373291,
+    "T->I_r10": 0.5625125169754028,
+    "T->I_r5": 0.5149030089378357
   },
   "_meta": {
     "audio_suffix": "mn20_audioheavy_lora1280_audio_features",
-    "checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v7_librispeech360_l4i/checkpoint_epoch_11.pt",
+    "checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v9_subjectfix_l4k/best_model.pt",
     "device": "NVIDIA GeForce GT 1030",
     "dims": [
       512
     ],
-    "encoder_name": "mobilenetv4_conv_medium",
-    "image_suffix": "mobilenetv4_conv_medium_image_features"
+    "encoder_name": "mobilenetv4_conv_medium.e180_r384_in12k",
+    "image_suffix": "mobilenetv4_conv_medium"
   },
   "speech_chatterbox-512": {
-    "A->T_r1": 0.46719998121261597,
-    "A->T_r10": 0.824999988079071,
-    "A->T_r5": 0.739799976348877,
-    "T->A_r1": 0.46059998869895935,
-    "T->A_r10": 0.8277999758720398,
-    "T->A_r5": 0.7425999641418457
+    "A->T_r1": 0.32760000228881836,
+    "A->T_r10": 0.717199981212616,
+    "A->T_r5": 0.6119999885559082,
+    "T->A_r1": 0.32019999623298645,
+    "T->A_r10": 0.7089999914169312,
+    "T->A_r5": 0.6046000123023987
   }
 }

subject_eval.json

@@ -1,118 +1,118 @@
 {
-  "checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v7_librispeech360_l4i/checkpoint_epoch_11.pt",
+  "checkpoint": "/shared/augmem/triembed/checkpoints/ess_aist_full_v9_subjectfix_l4k/best_model.pt",
   "split": "val",
-  "records_path": "/shared/augmem/triembed/checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v7_subject_media_wit4096_speech100k_wavcaps100k_librispeech360/val.jsonl",
+  "records_path": "/shared/augmem/triembed/checkpoints/ess_ait_86m_20260430T035907Z/ess_corpus_v9_subject_media_wit4096_speech100k_wavcaps100k_librispeech360_subjectfix/val.jsonl",
   "views": {
     "semantic_key": {
       "subject_same_different_auc": {
-        "auc": 0.4265240470767738,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7436165443684017,
-        "negative_mean": 0.7611866422867968
+        "auc": 0.9562558404233674,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.9163722881572546,
+        "negative_mean": 0.7671980599613585
       },
       "same_topic_different_subject_rejection_auc": {
-        "auc": 0.4265240470767738,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7436165443684017,
-        "negative_mean": 0.7611866422867968
+        "auc": 0.9562558404233674,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.9163722881572546,
+        "negative_mean": 0.7671980599613585
       }
     },
     "subject_key": {
       "subject_same_different_auc": {
-        "auc": 0.5066875746523821,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7964573271532047,
-        "negative_mean": 0.7948588548339001
+        "auc": 0.9881162919768391,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.8715863582124139,
+        "negative_mean": 0.507699292860264
       },
       "same_topic_different_subject_rejection_auc": {
-        "auc": 0.5066875746523821,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7964573271532047,
-        "negative_mean": 0.7948588548339001
+        "auc": 0.9881162919768391,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.8715863582124139,
+        "negative_mean": 0.507699292860264
       }
     },
     "event_key": {
       "subject_same_different_auc": {
-        "auc": 0.3832485276953712,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.6533856675037972,
-        "negative_mean": 0.7097943563909324
+        "auc": 0.9551271013544805,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.9246441395894791,
+        "negative_mean": 0.7915105798808771
       },
       "same_topic_different_subject_rejection_auc": {
-        "auc": 0.3832485276953712,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.6533856675037972,
-        "negative_mean": 0.7097943563909324
+        "auc": 0.9551271013544805,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.9246441395894791,
+        "negative_mean": 0.7915105798808771
       }
     },
     "full_key": {
       "subject_same_different_auc": {
-        "auc": 0.42067046032727157,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7333961783599913,
-        "negative_mean": 0.754728921120172
+        "auc": 0.9778614751548688,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.907507678522579,
+        "negative_mean": 0.7068911181573889
       },
       "same_topic_different_subject_rejection_auc": {
-        "auc": 0.42067046032727157,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7333961783599913,
-        "negative_mean": 0.754728921120172
+        "auc": 0.9778614751548688,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.907507678522579,
+        "negative_mean": 0.7068911181573889
       }
     },
     "prefix_512": {
       "subject_same_different_auc": {
-        "auc": 0.4265240470767738,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7436165443684017,
-        "negative_mean": 0.7611866422867968
+        "auc": 0.9562558404233674,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.9163722881572546,
+        "negative_mean": 0.7671980599613585
       },
       "same_topic_different_subject_rejection_auc": {
-        "auc": 0.4265240470767738,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7436165443684017,
-        "negative_mean": 0.7611866422867968
+        "auc": 0.9562558404233674,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.9163722881572546,
+        "negative_mean": 0.7671980599613585
       }
     },
     "prefix_1024": {
       "subject_same_different_auc": {
-        "auc": 0.4690923681629257,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7721689081039962,
-        "negative_mean": 0.7791595536982812
+        "auc": 0.9814636892484202,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.8989504970086771,
+        "negative_mean": 0.6642348380165352
       },
       "same_topic_different_subject_rejection_auc": {
-        "auc": 0.4690923681629257,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7721689081039962,
-        "negative_mean": 0.7791595536982812
+        "auc": 0.9814636892484202,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.8989504970086771,
+        "negative_mean": 0.6642348380165352
       }
     },
     "prefix_1536": {
       "subject_same_different_auc": {
-        "auc": 0.42067046032727157,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7333961783599913,
-        "negative_mean": 0.754728921120172
+        "auc": 0.9778614751548688,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.907507678522579,
+        "negative_mean": 0.7068911181573889
       },
       "same_topic_different_subject_rejection_auc": {
-        "auc": 0.42067046032727157,
-        "positive_pairs": 160248,
-        "negative_pairs": 6805,
-        "positive_mean": 0.7333961783599913,
-        "negative_mean": 0.754728921120172
+        "auc": 0.9778614751548688,
+        "positive_pairs": 119068,
+        "negative_pairs": 104655,
+        "positive_mean": 0.907507678522579,
+        "negative_mean": 0.7068911181573889
       }
     }
   }