| { | |
| "run_name": "rr_v2_rl_lite", | |
| "base_model": "Qwen/Qwen3-1.7B", | |
| "init_adapter": "outputs/research_reasoner_1p7b_v2_schema_hard", | |
| "output_dir": "outputs/research_reasoner_1p7b_v2_rl_lite", | |
| "max_seq_length": 4096, | |
| "packing": false, | |
| "lora": { | |
| "r": 32, | |
| "alpha": 64, | |
| "dropout": 0.05, | |
| "target_modules": [ | |
| "q_proj", | |
| "k_proj", | |
| "v_proj", | |
| "o_proj", | |
| "gate_proj", | |
| "up_proj", | |
| "down_proj" | |
| ] | |
| }, | |
| "optimizer": { | |
| "learning_rate": 5e-06, | |
| "warmup_ratio": 0.03, | |
| "weight_decay": 0.0, | |
| "lr_scheduler_type": "cosine", | |
| "optim": "paged_adamw_8bit" | |
| }, | |
| "batching": { | |
| "per_device_train_batch_size": 1, | |
| "gradient_accumulation_steps": 4, | |
| "num_train_epochs": 1 | |
| }, | |
| "logging": { | |
| "save_steps": 100, | |
| "eval_steps": 20, | |
| "logging_steps": 5, | |
| "report_to": "none" | |
| }, | |
| "data": { | |
| "train_file": "data/processed/rl_rejection_sft_v2_flat.jsonl", | |
| "eval_file": "data/processed/eval_v2_flat.jsonl" | |
| } | |
| } |