Add proof and artifacts table to model card
Browse files
README.md
CHANGED
|
@@ -154,6 +154,21 @@ The training curve and three-way before/after comparison are committed in the Gi
|
|
| 154 |
|
| 155 |
|
| 156 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 157 |
## 6) The honest trade-off
|
| 158 |
|
| 159 |
GRPO heavily reinforced the high-reward action patterns from training (Phase 2/3 episodes give +2.0 fix rewards vs Phase 1's +1.0 per violation). The trained model now over-applies these patterns to Phase 1 tasks where they don't fit, causing regressions on `validate_response_schema`, `validate_cross_field_constraints`, and `validate_auth_request`. With task-balanced training and a "don't repeat" reward signal, this would close. But the headroom-task win is real and reproducible.
|
|
|
|
| 154 |
|
| 155 |
|
| 156 |
|
| 157 |
+
### Proof and artifacts
|
| 158 |
+
|
| 159 |
+
The training run is backed by public, reproducible artifacts:
|
| 160 |
+
|
| 161 |
+
| Artifact | Link |
|
| 162 |
+
|---|---|
|
| 163 |
+
| Public WandB run | https://wandb.ai/pushpamsubscriptions-inn/openenv-contract-guardian/runs/gch0eg3k |
|
| 164 |
+
| Training proof summary | https://github.com/kumarpushpam17-personal/Hackathon/blob/main/api_contract_validator/results/TRAINING_RUN_PROOF.md |
|
| 165 |
+
| Full training log | https://github.com/kumarpushpam17-personal/Hackathon/blob/main/api_contract_validator/results/training_full_log.txt |
|
| 166 |
+
| Training state JSON | https://github.com/kumarpushpam17-personal/Hackathon/blob/main/api_contract_validator/results/training_state.json |
|
| 167 |
+
| Trained scores | https://github.com/kumarpushpam17-personal/Hackathon/blob/main/trained_scores.json |
|
| 168 |
+
| 7B baseline scores | https://github.com/kumarpushpam17-personal/Hackathon/blob/main/baseline_7b_scores.json |
|
| 169 |
+
| 72B baseline scores | https://github.com/kumarpushpam17-personal/Hackathon/blob/main/baseline_72b_v2_scores.json |
|
| 170 |
+
| Trained LoRA adapter | https://huggingface.co/pushpam14/api-contract-validator-grpo-7b |
|
| 171 |
+
|
| 172 |
## 6) The honest trade-off
|
| 173 |
|
| 174 |
GRPO heavily reinforced the high-reward action patterns from training (Phase 2/3 episodes give +2.0 fix rewards vs Phase 1's +1.0 per violation). The trained model now over-applies these patterns to Phase 1 tasks where they don't fit, causing regressions on `validate_response_schema`, `validate_cross_field_constraints`, and `validate_auth_request`. With task-balanced training and a "don't repeat" reward signal, this would close. But the headroom-task win is real and reproducible.
|