| { |
| "status": "ok", |
| "training_mode": "full", |
| "completed_models": 1, |
| "failed_or_skipped_models": 0, |
| "models": [ |
| { |
| "run_id": "qwen-qwen2-5-3b-instruct", |
| "training_mode": "full", |
| "model_id": "Qwen/Qwen2.5-3B-Instruct", |
| "label": "Qwen2.5-3B", |
| "status": "completed", |
| "error": "", |
| "sft_backend": "trl_transformers", |
| "sft_examples": 2000, |
| "sft_train_loss": 0.15688225453009363, |
| "sft_runtime": 715.2908, |
| "grpo_backend": "trl_transformers", |
| "grpo_records": 2000, |
| "grpo_avg_reward": 0.767, |
| "sft_inference_reward": 0.781, |
| "sft_valid_rate": 1.0, |
| "sft_latency_seconds": 2.863, |
| "grpo_inference_reward": 0.726, |
| "grpo_valid_rate": 1.0, |
| "grpo_latency_seconds": 3.681, |
| "train_holdout_gap": 0.041, |
| "fallback_detected": false, |
| "reward_range_ok": true, |
| "reward_range_failures": [], |
| "exploit_rate": 0.411, |
| "legal_rate": 0.93, |
| "candidate_diversity": 0.003, |
| "top_candidate_rate": 0.668, |
| "reward_components": { |
| "format_compliance_score": 0.999, |
| "candidate_alignment_score": 0.999, |
| "legality_score": 0.929, |
| "safety_delta_score": 0.497, |
| "burden_improvement_score": 0.469, |
| "disease_stability_score": 0.861, |
| "dosing_quality_score": 0.526, |
| "abstention_quality_score": 0.56, |
| "efficiency_score": 0.849, |
| "process_fidelity_score": 0.856, |
| "explanation_grounding_score": 0.795, |
| "anti_cheat_score": 0.589, |
| "uncertainty_calibration_score": 0.747 |
| }, |
| "primary_reward_channels": { |
| "safety_legality": 0.816, |
| "clinical_improvement": 0.609, |
| "dosing_quality": 0.543, |
| "process_integrity": 0.875 |
| }, |
| "artifact_paths": { |
| "sft": "/app/checkpoints/sweeps/qwen-qwen2-5-3b-instruct/sft_adapter", |
| "grpo": "/app/checkpoints/sweeps/qwen-qwen2-5-3b-instruct/grpo_adapter" |
| } |
| } |
| ], |
| "charts": { |
| "sft_vs_grpo_reward": "outputs/plots/sft_vs_grpo_reward.png", |
| "sft_loss_curves": "outputs/plots/sft_loss_curves.png", |
| "qwen_model_sft_reward": "outputs/plots/qwen_model_sft_reward.png", |
| "qwen_model_sft_loss": "outputs/plots/qwen_model_sft_loss.png", |
| "sft_validity_reward": "outputs/plots/sft_validity_reward.png", |
| "grpo_reward_curves": "outputs/plots/grpo_reward_curves.png", |
| "qwen_model_grpo_reward": "outputs/plots/qwen_model_grpo_reward.png", |
| "reward_component_bars": "outputs/plots/reward_component_bars.png", |
| "anti_cheat_failure_rates": "outputs/plots/anti_cheat_failure_rates.png", |
| "train_holdout_gap": "outputs/plots/train_holdout_gap.png", |
| "inference_validity_reward": "outputs/plots/inference_validity_reward.png", |
| "inference_latency_validity": "outputs/plots/inference_latency_validity.png" |
| } |
| } |
