Publish NexaMat rollout stack manifests and examples
Browse files- README.md +36 -0
- decoder/diffusion_v1/manifest.json +8 -0
- encoder/v1/manifest.json +8 -0
- eval/controller_smoke/all_predictions.jsonl +6 -0
- eval/controller_smoke/all_summary.json +32 -0
- eval/controller_smoke/reverse_predictions.jsonl +1 -0
- eval/controller_smoke/reverse_summary.json +32 -0
- examples/forward_stack_demo.json +22 -0
- examples/reverse_pool_demo.json +12 -0
- multimodal/controller/current/manifest.json +9 -0
- stack_manifest.json +121 -0
README.md
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: transformers
|
| 3 |
+
pipeline_tag: text-generation
|
| 4 |
+
tags:
|
| 5 |
+
- materials-science
|
| 6 |
+
- crystal-generation
|
| 7 |
+
- diffusion
|
| 8 |
+
- controller
|
| 9 |
+
- nexamat
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# NexaMat2
|
| 13 |
+
|
| 14 |
+
NexaMat2 is the rollout artifact hub for the NexaMat stack. It is not a single monolithic model.
|
| 15 |
+
It binds a frozen physics encoder, a constrained diffusion decoder, and a partially fine-tuned Qwen3-14B
|
| 16 |
+
controller into one auditable materials-discovery workflow.
|
| 17 |
+
|
| 18 |
+
## Current Controller
|
| 19 |
+
|
| 20 |
+
```text
|
| 21 |
+
model_id: AethronPhantom/Nexa_Mat2
|
| 22 |
+
subfolder: multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged
|
| 23 |
+
base_model: Qwen/Qwen3-14B
|
| 24 |
+
state: partial FFT pilot, not full-dataset FFT
|
| 25 |
+
```
|
| 26 |
+
|
| 27 |
+
## Stack Contract
|
| 28 |
+
|
| 29 |
+
- Encoder: physics grounding and plausibility scoring.
|
| 30 |
+
- Decoder: best-of-N constrained candidate proposal.
|
| 31 |
+
- Controller: ranking, rejection, evidence routing, and explanation.
|
| 32 |
+
- Cross-attention bridge: adapter surface for text, encoder, decoder, task, and evidence streams.
|
| 33 |
+
|
| 34 |
+
See `stack_manifest.json` for the machine-readable rollout contract.
|
| 35 |
+
|
| 36 |
+
Status: `rollout_partial_controller`.
|
decoder/diffusion_v1/manifest.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"artifact_uri": "wasabi:nexa-mat/models/nexa_mat/diffusion-generation-v1/prod/nexa_mat_diffusion_prod_fp32_b512_phase_sampler_20260512T162654Z/final_checkpoint.pt",
|
| 3 |
+
"closeout_report": "docs/projects/Nexa_Mat/results/reports/2026-05-13_diffusion_prod_225k_closeout.md",
|
| 4 |
+
"generation_mode": "distributional_sampler_best_of_n",
|
| 5 |
+
"name": "nexa_mat_diffusion_generation_v1",
|
| 6 |
+
"role": "constrained best-of-N candidate proposal and structure repair",
|
| 7 |
+
"run_root_uri": "wasabi:nexa-mat/models/nexa_mat/diffusion-generation-v1/prod/nexa_mat_diffusion_prod_fp32_b512_phase_sampler_20260512T162654Z"
|
| 8 |
+
}
|
encoder/v1/manifest.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"artifact_uri": "wasabi:nexa-mat/models/nexa_mat/encoder-pretrain/full_v1/2026-05-06/nexa_mat_encoder_blackwell_full_v1_continue_to_200k_20260506T075034Z/nexa_mat_V1_final.pt",
|
| 3 |
+
"closeout_report": "docs/projects/Nexa_Mat/results/reports/2026-05-06_full_pretrain_v1_200k_closeout.md",
|
| 4 |
+
"dataset_uri": "wasabi:nexa-mat/datasets/nexa_mat_v1",
|
| 5 |
+
"frozen": true,
|
| 6 |
+
"name": "nexa_mat_V1_final",
|
| 7 |
+
"role": "frozen physics grounding and plausibility scoring"
|
| 8 |
+
}
|
eval/controller_smoke/all_predictions.jsonl
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"record_id": "af23cbb7e220e94da7a4d5c0", "prediction": "No markdown, no text, no citations, no explanations. Return structured decision with constraints and candidate routing.\nTARGET_DECISION:\n{\"candidate_decisions\":[{\"candidate_id\":\"cand_frontier_001\",\"decision_type\":\"frontier_dft_candidate\",\"requires_followup\":[\"full_dft_validation\",\"structure_or_property_evidence\"],\"usable_for\":[\"candidate_ranking\",\"evidence_planning\",\"dft_planning\"]}],\"decision_type\":\"stack_agreement_decision\",\"requires_followup\":[\"full_dft_validation\",\"structure_or_property_evidence\"],\"usable_for\":[\"candidate_ranking\",\"evidence_planning\",\"dft_planning\"]}", "generator": "AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"}
|
| 2 |
+
{"record_id": "9987417195968e8d6910f9c7", "prediction": "No markdown, no text, no explanations. Return a structured decision with constraints, candidate routing, reject reasons, and the next action for evidence retrieval or human review.\nTARGET_DECISION:\n{\"decision_type\":\"stack_agreement_decision\",\"decision_role\":\"candidate_router_or_retrieval_planner\",\"requires_followup\":[\"full_dft_or_synthesis_validation\",\"structure_or_property_evidence_check\"],\"usable_for\":[\"candidate_ranking\",\"evidence_planning\",\"decision_validation\"],\"venue\":\"materials_discovery_or_synthesis_planning\",\"workflow_stage\":\"preliminary_screening_or_planning\"}", "generator": "AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"}
|
| 3 |
+
{"record_id": "d31ab043d82d72e120834a7c", "prediction": "No markdown, no text, no citations, no explanations. Return structured decision with constraints, candidate routing, reject reasons, evidence references, and the next action for decoder sampling, evidence retrieval, DFT validation, or human review.\nTARGET_DECISION:\n{\"candidate_decisions\":[{\"candidate_id\":\"cand_ranked_001\",\"decision_type\":\"stack_agreement_decision\",\"requires_followup\":[\"full_dft_validation\"],\"usable_for\":[\"candidate_ranking\",\"decision_grounding\"],\"venue_context\":null,\"was_retrieved_for\":[\"decision_grounding\",\"evidence_retrieval_context\"]}],\"decision_type\":\"candidate_ranking_or_retrieval_context\",\"requires_followup\":[\"full_dft_validation\"],\"usable_for\":[\"candidate_ranking\",\"decision_grounding\",\"evidence_retrieval_context\"],\"venue_context\":null,\"was_retrieved_for\":[\"decision_grounding\",\"evidence_retrieval_context\"]}", "generator": "AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"}
|
| 4 |
+
{"record_id": "0908c8989f43dcc42f5058eb", "prediction": "No explanations or markdown.\n{\"decision_type\":\"stack_agreement_decision\",\"decision_reason\":\"Physically invalid candidate with hard constraint violations (min_distance_angstrom=0.48, min_distance_violation_rate=0.12, passes_basic_constraints=false, passes_plausibility_filter=false)\",\"candidate_id\":\"cand_invalid_001\",\"decision_class\":\"rejected_candidates\",\"requires_followup\":[\"full_geometry_check\",\"structure_repair_or_rejection\"],\"not_safe_for\":[\"evidence_retrieval\",\"candidate_promotion\"]}", "generator": "AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"}
|
| 5 |
+
{"record_id": "bcc8353b05fa8714b9222c43", "prediction": "No markdown, no text, no citations, no explanations. Return structured constraints and evidence references only. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate_id, decision_type, decision_reason, evidence_references, and candidate_pool fields. Return null for any fields not explicitly provided in the input. Return the candidate pool as-is with no modifications. Return the structured decision with constraints, evidence references, and the next action for each candidate. Return only the candidate", "generator": "AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"}
|
| 6 |
+
{"record_id": "42d2dff376f8fb468da28fd2", "prediction": "No markdown, no text, no citations, no explanations. Return structured decision with constraints, candidate routing, reject reasons, evidence references, and the next action for each candidate in the pool.\nTARGET_DECISION:\n{\"candidate_decisions\":[{\"candidate_id\":\"cand_conservative_known_001\",\"decision_type\":\"stack_agreement_decision\",\"decision_role\":\"candidate_router_or_retrieval_filter\",\"requires_followup\":[\"full_dft_or_synthesis_validation\"],\"usable_for\":[\"candidate_retrieval_filtering\",\"decision_stack_grounding\",\"evidence_context_provision\"],\"venue_context\":\"materials_synthesis_or_property_prediction\",\"was_retrieved_from\":[\"candidate_pool\",\"decision_stack_grounding\",\"evidence_context_provision\"],\"will_be_used_for\":[\"candidate_retrieval_filtering\",\"decision_stack_grounding\",\"evidence_context_provision\"]}],\"decision_type\":\"stack_agreement_decision\",\"decision_role\":\"candidate_router_or_retrieval_filter\",\"requires_followup\":[\"full_dft_or_synthesis_validation\"],\"usable_for\":[\"candidate_retrieval_filtering\",\"decision_stack_grounding\",\"evidence_context_provision\"],\"venue_context\":\"materials_synthesis_or_property_prediction\",\"was_retrieved_from\":[\"candidate_pool\",\"decision_stack_grounding\",\"evidence_context_provision\"],\"will_be_used_for\":[\"candidate_retrieval_filtering\",\"decision_stack_grounding\",\"evidence_context_provision\"]}", "generator": "AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"}
|
eval/controller_smoke/all_summary.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alignment_gates": 3,
|
| 3 |
+
"alignment_score": 0.4315514275885793,
|
| 4 |
+
"candidate_decision_accuracy": 0.16666666666666666,
|
| 5 |
+
"candidates": 6,
|
| 6 |
+
"controller_alignment_promotable": false,
|
| 7 |
+
"criteria": {
|
| 8 |
+
"alignment_ready_for_fft": true,
|
| 9 |
+
"bad_candidate_false_positive_max": 0.0,
|
| 10 |
+
"min_frontier_preservation": 1.0,
|
| 11 |
+
"min_known_deprioritization": 1.0,
|
| 12 |
+
"uncontained_hallucination_max": 0.0
|
| 13 |
+
},
|
| 14 |
+
"false_positive_rate_on_bad_candidates": 0.0,
|
| 15 |
+
"frontier_preservation_rate": 1.0,
|
| 16 |
+
"gate_checks": {
|
| 17 |
+
"bad_candidate_false_positive": true,
|
| 18 |
+
"frontier_preservation": true,
|
| 19 |
+
"known_deprioritization": false,
|
| 20 |
+
"ready_for_fft": false,
|
| 21 |
+
"uncontained_hallucination": false
|
| 22 |
+
},
|
| 23 |
+
"known_deprioritization_rate": 0.0,
|
| 24 |
+
"prediction_generators": [
|
| 25 |
+
"AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"
|
| 26 |
+
],
|
| 27 |
+
"ready_for_fft": false,
|
| 28 |
+
"records": 6,
|
| 29 |
+
"report": "/dev/shm/nexa_mat_controller_rollout_all_eval.md",
|
| 30 |
+
"structured_output_validity": 0.16666666666666666,
|
| 31 |
+
"uncontained_hallucination_rate": 1.0
|
| 32 |
+
}
|
eval/controller_smoke/reverse_predictions.jsonl
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"record_id": "171178b159d3c16c09173074", "prediction": "No markdown, no text, no citations, no explanations. Return structured decisions with constraints and evidence references.\nTARGET_DECISION:\n{\"decision_type\":\"stack_agreement_decision\",\"evidence_role\":\"candidate_pool_ranking_or_retrieval_filtering\",\"requires_followup\":[\"full_dft_validation\",\"structure_or_property_retrieval\"],\"usable_for\":[\"candidate_pool_ranking\",\"retrieval_filtering\"],\"valid_for\":[\"materials_synthesis_planning\",\"structure_or_property_retrieval\",\"candidate_pool_ranking\"],\"candidate_decisions\":[{\"candidate_id\":\"cand_frontier_001\",\"decision_type\":\"frontier_dft_candidate\",\"requires_followup\":[\"full_dft_validation\"],\"usable_for\":[\"materials_synthesis_planning\"],\"valid_for\":[\"materials_synthesis_planning\"],\"evidence_role\":\"candidate_pool_ranking_or_retrieval_filtering\"},{\"candidate_id\":\"cand_known_001\",\"decision_type\":\"known_deprioritized\",\"requires_followup\":[\"structure_or_property_retrieval\"],\"usable_for\":[\"candidate_pool_ranking\"],\"valid_for\":[\"candidate_pool_ranking\"],\"evidence_role\":\"candidate_pool_ranking_or_retrieval_filtering\"},{\"candidate_id\":\"cand_ranked_001\",\"decision_type\":\"ranked_candidate\",\"requires_followup\":[\"structure_or_property_retrieval\"],\"usable_for\":[\"candidate_pool_ranking\",\"retrieval_filtering\"],\"valid_for\":[\"candidate_pool_ranking\",\"retrieval_filtering\"],\"evidence_role\":\"candidate_pool_ranking_or_retrieval_filtering\"},{\"candidate_id\":\"cand_invalid_001\",\"decision_type\":\"reject_invalid\",\"requires_followup\":[\"structure_or_property_retrieval\"],\"usable_for\":[\"candidate_pool_ranking\"],\"valid_for\":[\"candidate_pool_ranking\"],\"evidence_role\":\"candidate_pool_ranking_or_retrieval_filtering\"},{\"candidate_id\":\"cand_retrieve_more_001\",\"decision_type\":\"retrieve_more\",\"requires_followup\":[\"structure_or_property_retrieval\"],\"usable_for\":[\"candidate_pool_ranking\",\"retrieval_filtering\"],\"valid_for\":[\"candidate_pool_ranking\",\"retrieval_filtering\"],\"evidence_role\":\"candidate_pool_ranking_or_retrieval_filtering\"},{\"candidate_id\":\"cand_conservative_known_001\",\"decision_type\":\"ranked_candidate\",\"requires_followup\":[\"structure_or_property_retrieval\"],\"usable_for\":[\"candidate_pool_ranking\",\"retrieval_filtering\"],\"valid_for\":[\"candidate_pool_ranking\",\"retrieval_filtering\"],\"evidence_role\":\"candidate_pool_ranking_or_retrieval_filtering\"}]}", "generator": "AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"}
|
eval/controller_smoke/reverse_summary.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alignment_gates": 10,
|
| 3 |
+
"alignment_score": 1.0,
|
| 4 |
+
"candidate_decision_accuracy": 1.0,
|
| 5 |
+
"candidates": 6,
|
| 6 |
+
"controller_alignment_promotable": true,
|
| 7 |
+
"criteria": {
|
| 8 |
+
"alignment_ready_for_fft": true,
|
| 9 |
+
"bad_candidate_false_positive_max": 0.0,
|
| 10 |
+
"min_frontier_preservation": 1.0,
|
| 11 |
+
"min_known_deprioritization": 1.0,
|
| 12 |
+
"uncontained_hallucination_max": 0.0
|
| 13 |
+
},
|
| 14 |
+
"false_positive_rate_on_bad_candidates": 0.0,
|
| 15 |
+
"frontier_preservation_rate": 1.0,
|
| 16 |
+
"gate_checks": {
|
| 17 |
+
"bad_candidate_false_positive": true,
|
| 18 |
+
"frontier_preservation": true,
|
| 19 |
+
"known_deprioritization": true,
|
| 20 |
+
"ready_for_fft": true,
|
| 21 |
+
"uncontained_hallucination": true
|
| 22 |
+
},
|
| 23 |
+
"known_deprioritization_rate": 1.0,
|
| 24 |
+
"prediction_generators": [
|
| 25 |
+
"AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"
|
| 26 |
+
],
|
| 27 |
+
"ready_for_fft": true,
|
| 28 |
+
"records": 1,
|
| 29 |
+
"report": "/dev/shm/nexa_mat_controller_rollout_reverse_eval.md",
|
| 30 |
+
"structured_output_validity": 1.0,
|
| 31 |
+
"uncontained_hallucination_rate": 0.0
|
| 32 |
+
}
|
examples/forward_stack_demo.json
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"mode": "forward",
|
| 3 |
+
"sample_count": 6,
|
| 4 |
+
"samples": [
|
| 5 |
+
{
|
| 6 |
+
"generator": "rule_based_stack_controller",
|
| 7 |
+
"prediction": "{\"decision_type\":\"stack_agreement_decision\",\"deprioritized_candidates\":[],\"dft_candidates\":[{\"candidate_id\":\"cand_frontier_001\",\"reason\":\"frontier_dft_candidate\"}],\"explanation\":\"Deterministic stack-policy baseline using validators, novelty, knownness, and evidence containment.\",\"next_action\":\"rank_and_route_candidate_pool\",\"ranked_candidates\":[],\"rejected_candidates\":[],\"retrieve_more_candidates\":[]}",
|
| 8 |
+
"record_id": "af23cbb7e220e94da7a4d5c0"
|
| 9 |
+
},
|
| 10 |
+
{
|
| 11 |
+
"generator": "rule_based_stack_controller",
|
| 12 |
+
"prediction": "{\"decision_type\":\"stack_agreement_decision\",\"deprioritized_candidates\":[{\"candidate_id\":\"cand_known_001\",\"reason\":\"known_deprioritized\"}],\"dft_candidates\":[],\"explanation\":\"Deterministic stack-policy baseline using validators, novelty, knownness, and evidence containment.\",\"next_action\":\"rank_and_route_candidate_pool\",\"ranked_candidates\":[],\"rejected_candidates\":[],\"retrieve_more_candidates\":[]}",
|
| 13 |
+
"record_id": "9987417195968e8d6910f9c7"
|
| 14 |
+
},
|
| 15 |
+
{
|
| 16 |
+
"generator": "rule_based_stack_controller",
|
| 17 |
+
"prediction": "{\"decision_type\":\"stack_agreement_decision\",\"deprioritized_candidates\":[],\"dft_candidates\":[],\"explanation\":\"Deterministic stack-policy baseline using validators, novelty, knownness, and evidence containment.\",\"next_action\":\"rank_and_route_candidate_pool\",\"ranked_candidates\":[{\"candidate_id\":\"cand_ranked_001\",\"reason\":\"ranked_candidate\"}],\"rejected_candidates\":[],\"retrieve_more_candidates\":[]}",
|
| 18 |
+
"record_id": "d31ab043d82d72e120834a7c"
|
| 19 |
+
}
|
| 20 |
+
],
|
| 21 |
+
"source": "docs/projects/Nexa_Mat/results/reports/stack_agreement_fixture_20260518/forward_rule_based_predictions.jsonl"
|
| 22 |
+
}
|
examples/reverse_pool_demo.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"mode": "reverse",
|
| 3 |
+
"sample_count": 1,
|
| 4 |
+
"samples": [
|
| 5 |
+
{
|
| 6 |
+
"generator": "rule_based_stack_controller",
|
| 7 |
+
"prediction": "{\"decision_type\":\"stack_agreement_decision\",\"deprioritized_candidates\":[{\"candidate_id\":\"cand_known_001\",\"reason\":\"known_deprioritized\"}],\"dft_candidates\":[{\"candidate_id\":\"cand_frontier_001\",\"reason\":\"frontier_dft_candidate\"}],\"explanation\":\"Deterministic stack-policy baseline using validators, novelty, knownness, and evidence containment.\",\"next_action\":\"rank_and_route_candidate_pool\",\"ranked_candidates\":[{\"candidate_id\":\"cand_ranked_001\",\"reason\":\"ranked_candidate\"},{\"candidate_id\":\"cand_conservative_known_001\",\"reason\":\"ranked_candidate\"}],\"rejected_candidates\":[{\"candidate_id\":\"cand_invalid_001\",\"reason\":\"reject_invalid\"}],\"retrieve_more_candidates\":[{\"candidate_id\":\"cand_retrieve_more_001\",\"reason\":\"retrieve_more\"}]}",
|
| 8 |
+
"record_id": "171178b159d3c16c09173074"
|
| 9 |
+
}
|
| 10 |
+
],
|
| 11 |
+
"source": "docs/projects/Nexa_Mat/results/reports/stack_agreement_fixture_20260518/reverse_pool_rule_based_predictions.jsonl"
|
| 12 |
+
}
|
multimodal/controller/current/manifest.json
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"base_model": "Qwen/Qwen3-14B",
|
| 3 |
+
"closeout_report": "docs/projects/Nexa_Mat/results/reports/2026-05-19_multimodal_h200_fft_closeout.md",
|
| 4 |
+
"model_repo": "AethronPhantom/Nexa_Mat2",
|
| 5 |
+
"model_subfolder": "multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged",
|
| 6 |
+
"name": "nexa_mat_controller_qwen3_14b_fft_pilot",
|
| 7 |
+
"role": "semantic controller for ranking, rejection, evidence routing, and explanation",
|
| 8 |
+
"training_state": "partial_fft_pilot_not_full_dataset"
|
| 9 |
+
}
|
stack_manifest.json
ADDED
|
@@ -0,0 +1,121 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"components": {
|
| 3 |
+
"controller": {
|
| 4 |
+
"base_model": "Qwen/Qwen3-14B",
|
| 5 |
+
"closeout_report": "docs/projects/Nexa_Mat/results/reports/2026-05-19_multimodal_h200_fft_closeout.md",
|
| 6 |
+
"model_repo": "AethronPhantom/Nexa_Mat2",
|
| 7 |
+
"model_subfolder": "multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged",
|
| 8 |
+
"name": "nexa_mat_controller_qwen3_14b_fft_pilot",
|
| 9 |
+
"role": "semantic controller for ranking, rejection, evidence routing, and explanation",
|
| 10 |
+
"training_state": "partial_fft_pilot_not_full_dataset"
|
| 11 |
+
},
|
| 12 |
+
"cross_attention_bridge": {
|
| 13 |
+
"module": "nexa_train.materials.cross_attention.GatedCrossAttentionBridge",
|
| 14 |
+
"role": "inject text, encoder, decoder, task, and evidence streams into diffusion/controller memory",
|
| 15 |
+
"status": "implemented_adapter_surface"
|
| 16 |
+
},
|
| 17 |
+
"decoder": {
|
| 18 |
+
"artifact_uri": "wasabi:nexa-mat/models/nexa_mat/diffusion-generation-v1/prod/nexa_mat_diffusion_prod_fp32_b512_phase_sampler_20260512T162654Z/final_checkpoint.pt",
|
| 19 |
+
"closeout_report": "docs/projects/Nexa_Mat/results/reports/2026-05-13_diffusion_prod_225k_closeout.md",
|
| 20 |
+
"generation_mode": "distributional_sampler_best_of_n",
|
| 21 |
+
"name": "nexa_mat_diffusion_generation_v1",
|
| 22 |
+
"role": "constrained best-of-N candidate proposal and structure repair",
|
| 23 |
+
"run_root_uri": "wasabi:nexa-mat/models/nexa_mat/diffusion-generation-v1/prod/nexa_mat_diffusion_prod_fp32_b512_phase_sampler_20260512T162654Z"
|
| 24 |
+
},
|
| 25 |
+
"encoder": {
|
| 26 |
+
"artifact_uri": "wasabi:nexa-mat/models/nexa_mat/encoder-pretrain/full_v1/2026-05-06/nexa_mat_encoder_blackwell_full_v1_continue_to_200k_20260506T075034Z/nexa_mat_V1_final.pt",
|
| 27 |
+
"closeout_report": "docs/projects/Nexa_Mat/results/reports/2026-05-06_full_pretrain_v1_200k_closeout.md",
|
| 28 |
+
"dataset_uri": "wasabi:nexa-mat/datasets/nexa_mat_v1",
|
| 29 |
+
"frozen": true,
|
| 30 |
+
"name": "nexa_mat_V1_final",
|
| 31 |
+
"role": "frozen physics grounding and plausibility scoring"
|
| 32 |
+
}
|
| 33 |
+
},
|
| 34 |
+
"created_at_utc": "2026-05-19T21:53:54.250964+00:00",
|
| 35 |
+
"intended_use": "forward candidate proposal/ranking and reverse candidate-pool triage",
|
| 36 |
+
"learned_controller_smoke": {
|
| 37 |
+
"all_fixture": {
|
| 38 |
+
"alignment_gates": 3,
|
| 39 |
+
"alignment_score": 0.4315514275885793,
|
| 40 |
+
"candidate_decision_accuracy": 0.16666666666666666,
|
| 41 |
+
"candidates": 6,
|
| 42 |
+
"controller_alignment_promotable": false,
|
| 43 |
+
"criteria": {
|
| 44 |
+
"alignment_ready_for_fft": true,
|
| 45 |
+
"bad_candidate_false_positive_max": 0.0,
|
| 46 |
+
"min_frontier_preservation": 1.0,
|
| 47 |
+
"min_known_deprioritization": 1.0,
|
| 48 |
+
"uncontained_hallucination_max": 0.0
|
| 49 |
+
},
|
| 50 |
+
"false_positive_rate_on_bad_candidates": 0.0,
|
| 51 |
+
"frontier_preservation_rate": 1.0,
|
| 52 |
+
"gate_checks": {
|
| 53 |
+
"bad_candidate_false_positive": true,
|
| 54 |
+
"frontier_preservation": true,
|
| 55 |
+
"known_deprioritization": false,
|
| 56 |
+
"ready_for_fft": false,
|
| 57 |
+
"uncontained_hallucination": false
|
| 58 |
+
},
|
| 59 |
+
"known_deprioritization_rate": 0.0,
|
| 60 |
+
"prediction_generators": [
|
| 61 |
+
"AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"
|
| 62 |
+
],
|
| 63 |
+
"ready_for_fft": false,
|
| 64 |
+
"records": 6,
|
| 65 |
+
"report": "/dev/shm/nexa_mat_controller_rollout_all_eval.md",
|
| 66 |
+
"structured_output_validity": 0.16666666666666666,
|
| 67 |
+
"uncontained_hallucination_rate": 1.0
|
| 68 |
+
},
|
| 69 |
+
"artifact_paths": {
|
| 70 |
+
"all_predictions.jsonl": "eval/controller_smoke/all_predictions.jsonl",
|
| 71 |
+
"all_summary.json": "eval/controller_smoke/all_summary.json",
|
| 72 |
+
"reverse_predictions.jsonl": "eval/controller_smoke/reverse_predictions.jsonl",
|
| 73 |
+
"reverse_summary.json": "eval/controller_smoke/reverse_summary.json"
|
| 74 |
+
},
|
| 75 |
+
"read": "The partial FFT controller reloads and passes the reverse-pool routing smoke with 1024 generation tokens. It does not yet pass the broader all-fixture learned-controller gate, so rollout defaults should keep the deterministic stack policy as the conservative public path.",
|
| 76 |
+
"reverse_pool": {
|
| 77 |
+
"alignment_gates": 10,
|
| 78 |
+
"alignment_score": 1.0,
|
| 79 |
+
"candidate_decision_accuracy": 1.0,
|
| 80 |
+
"candidates": 6,
|
| 81 |
+
"controller_alignment_promotable": true,
|
| 82 |
+
"criteria": {
|
| 83 |
+
"alignment_ready_for_fft": true,
|
| 84 |
+
"bad_candidate_false_positive_max": 0.0,
|
| 85 |
+
"min_frontier_preservation": 1.0,
|
| 86 |
+
"min_known_deprioritization": 1.0,
|
| 87 |
+
"uncontained_hallucination_max": 0.0
|
| 88 |
+
},
|
| 89 |
+
"false_positive_rate_on_bad_candidates": 0.0,
|
| 90 |
+
"frontier_preservation_rate": 1.0,
|
| 91 |
+
"gate_checks": {
|
| 92 |
+
"bad_candidate_false_positive": true,
|
| 93 |
+
"frontier_preservation": true,
|
| 94 |
+
"known_deprioritization": true,
|
| 95 |
+
"ready_for_fft": true,
|
| 96 |
+
"uncontained_hallucination": true
|
| 97 |
+
},
|
| 98 |
+
"known_deprioritization_rate": 1.0,
|
| 99 |
+
"prediction_generators": [
|
| 100 |
+
"AethronPhantom/Nexa_Mat2/multimodal/controller/nexa_mat_controller_fft_pilot_20260518T234148Z/final_model_merged"
|
| 101 |
+
],
|
| 102 |
+
"ready_for_fft": true,
|
| 103 |
+
"records": 1,
|
| 104 |
+
"report": "/dev/shm/nexa_mat_controller_rollout_reverse_eval.md",
|
| 105 |
+
"structured_output_validity": 1.0,
|
| 106 |
+
"uncontained_hallucination_rate": 0.0
|
| 107 |
+
}
|
| 108 |
+
},
|
| 109 |
+
"routing_contract": {
|
| 110 |
+
"forward": "intent -> constraints -> decoder samples -> encoder/validator filters -> controller ranks/explains",
|
| 111 |
+
"hard_rules": [
|
| 112 |
+
"reject hard physics failures",
|
| 113 |
+
"preserve physically plausible low-evidence novelty as DFT candidates",
|
| 114 |
+
"deprioritize known duplicates unless conservative mode allows them",
|
| 115 |
+
"route unsupported claims to retrieve-more or human review"
|
| 116 |
+
],
|
| 117 |
+
"reverse": "candidate pool -> encoder/task/evidence scoring -> optional decoder expansion -> controller rank/filter/explain"
|
| 118 |
+
},
|
| 119 |
+
"schema_version": "nexa_mat_stack_manifest/v1",
|
| 120 |
+
"status": "rollout_partial_controller"
|
| 121 |
+
}
|