{
  "mode": "train",
  "lineage": {
    "created_at_utc": "2026-04-12T22:14:08Z",
    "run_id": "shaer_grpo_20260412_221146",
    "run_dir": "/root/workspace/Shaer/grpo/outputs/train/shaer_grpo_20260412_221146",
    "chain_id": "shaer_grpo_20260412_221146",
    "root_run_id": "shaer_grpo_20260412_221146",
    "parent_run_id": "",
    "parent_run_dir": "",
    "run_sequence_index": 0
  },
  "dataset": {
    "dataset_id": "Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1",
    "dataset_source_id": "Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500-splits",
    "train_split": "train",
    "eval_split": "eval",
    "test_split": "test",
    "train_size": 9955,
    "eval_size": 80,
    "test_size": 80,
    "hard_diagnostic_size": 3126,
    "phase1_max_bayts": "20",
    "allowed_meters": [],
    "train_manifest_path": "/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_short_drop_trio/cap_1000/selected_manifest.csv",
    "hard_diagnostic_manifest_path": "/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_short_drop_trio/hard_diagnostic_cap_256/selected_manifest.csv"
  },
  "rewards": {
    "active_rewards": [
      "meter",
      "count_adherence",
      "hard_gate",
      "judge_quality",
      "repeat_soft",
      "total_composite"
    ],
    "weights": {
      "meter": 0.0,
      "count_adherence": 0.0,
      "hard_gate": 0.0,
      "judge_quality": 0.0,
      "repeat_soft": 0.0,
      "total_composite": 1.0
    }
  },
  "model": {
    "base_model_id": "Navid-AI/Yehia-7B-preview",
    "sft_adapter_repo": "Shaer-AI/Shaer-adapters",
    "sft_adapter_mode": "fresh_sft/train",
    "load_in_4bit_requested": true,
    "gradient_checkpointing": true
  },
  "effective_generation": {
    "use_vllm": true,
    "vllm_mode": "colocate",
    "vllm_gpu_memory_utilization": 0.5,
    "max_prompt_length": 1024,
    "max_completion_length": 512,
    "temperature": 0.9,
    "top_p": 1.0,
    "num_generations": 8,
    "num_generations_eval": 1
  },
  "effective_trainer": {
    "learning_rate": 1e-05,
    "per_device_train_batch_size": 1,
    "per_device_eval_batch_size_configured": 8,
    "per_device_eval_batch_size": 8,
    "gradient_accumulation_steps": 8,
    "max_steps": 3300,
    "logging_steps": 1,
    "eval_steps": 50,
    "save_steps": 50,
    "save_total_limit": 4,
    "beta": 0.0,
    "scale_rewards": "group",
    "loss_type": "dapo",
    "mask_truncated_completions": true,
    "bf16": true,
    "load_best_model_at_end": true,
    "metric_for_best_model": "eval_reward_total_mean",
    "greater_is_better": true,
    "hub_strategy": "HubStrategy.CHECKPOINT",
    "push_to_hub": true,
    "hub_model_id": "Shaer-AI/Shaer-adapters-grpo-short1k-no-trio-v2",
    "generations_repo_id": "Shaer-AI/Shaer-adapters-grpo-short1k-no-trio-v2-generations"
  },
  "sanity_mode_note": "When mode='sanity', these effective values override the base trainer/generation blocks from config_snapshot.json."
}