| { |
| "run": { |
| "seed": 42, |
| "run_name_prefix": "shaer_grpo", |
| "output_root": "outputs/train", |
| "base_model_id": "Navid-AI/Yehia-7B-preview", |
| "start_adapter_repo": "Shaer-AI/Shaer-adapters", |
| "start_adapter_mode": "fresh_sft/train", |
| "output_model_repo": "Shaer-AI/Shaer-adapters-grpo" |
| }, |
| "dataset": { |
| "train_dataset_id": "Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1", |
| "source_dataset_id": "Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500-splits", |
| "train_manifest_path": "outputs/curated_meter_count_locked/cap_3000/selected_manifest.csv", |
| "hard_diagnostic_manifest_path": "outputs/curated_meter_count_locked/hard_diagnostic_cap_256/selected_manifest.csv", |
| "train_split": "train", |
| "eval_split": "eval", |
| "test_split": "test", |
| "eval_bank_per_meter_per_bucket": 2, |
| "test_bank_per_meter_per_bucket": 4 |
| }, |
| "studies": { |
| "inspect_lengths": { |
| "output_root": "outputs/inspect_lengths", |
| "coverage_target": 0.9, |
| "push_filtered_default": false |
| }, |
| "publish_grpo_splits": { |
| "output_root": "outputs/publish_grpo_splits", |
| "seed": 42, |
| "validation_total": 104, |
| "test_total": 208, |
| "validation_short_per_meter": 4, |
| "validation_full_per_meter": 4, |
| "test_short_per_meter": 8, |
| "test_full_per_meter": 8 |
| }, |
| "sanity_check_model": { |
| "output_root": "outputs/sanity_check_model_base_n5_k5", |
| "backend": "vllm", |
| "n_prompts_per_meter": 5, |
| "k_generations": 5, |
| "thresholds": [ |
| 0.3, |
| 0.5, |
| 0.7, |
| 0.9 |
| ], |
| "max_new_tokens": 384, |
| "temperature": 0.95, |
| "top_p": 0.95, |
| "seed": 42, |
| "vllm_gpu_memory_utilization": 0.3, |
| "vllm_tensor_parallel_size": 1, |
| "vllm_dtype": "bfloat16", |
| "vllm_max_loras": 1, |
| "vllm_max_lora_rank": 64, |
| "flashinfer_disable_version_check": true |
| }, |
| "sanity_check_rewards": { |
| "output_root": "outputs/sanity_check_rewards", |
| "golden_samples": 39, |
| "seed": 42, |
| "sample_strategy": "balanced_by_base_meter" |
| } |
| }, |
| "phase1": { |
| "active_rewards": [ |
| "meter", |
| "count_adherence", |
| "arabic_clean", |
| "hard_gate", |
| "arabic_floor", |
| "count_floor", |
| "lexical_plausibility", |
| "judge_quality", |
| "repeat_penalty", |
| "repeat_floor", |
| "near_duplicate_penalty", |
| "opening_diversity", |
| "distinct_2", |
| "total_composite" |
| ], |
| "inactive_rewards": [ |
| "exact_count_bonus", |
| "meter_count_clean", |
| "meaning_fit", |
| "meaning_substance", |
| "fluency", |
| "poeticness", |
| "cohesion" |
| ], |
| "reward_weights": { |
| "meter": 0.0, |
| "count_adherence": 0.0, |
| "arabic_clean": 0.0, |
| "hard_gate": 0.0, |
| "arabic_floor": 0.0, |
| "count_floor": 0.0, |
| "lexical_plausibility": 0.0, |
| "judge_quality": 0.0, |
| "repeat_penalty": 0.0, |
| "repeat_floor": 0.0, |
| "near_duplicate_penalty": 0.0, |
| "opening_diversity": 0.0, |
| "distinct_2": 0.0, |
| "total_composite": 1.0 |
| }, |
| "judge_quality_prompt_file": "prompts/judge_quality_selected.yaml", |
| "judge_quality_cache_dir": "outputs/judge_quality_cache", |
| "judge_quality_max_workers": 4 |
| }, |
| "model": { |
| "load_in_4bit": true, |
| "bnb_4bit_quant_type": "nf4", |
| "bnb_4bit_use_double_quant": true, |
| "gradient_checkpointing": true |
| }, |
| "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 |
| }, |
| "trainer": { |
| "learning_rate": 1e-05, |
| "per_device_train_batch_size": 1, |
| "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, |
| "remove_unused_columns": false, |
| "report_to": [], |
| "bf16": true, |
| "load_best_model_at_end": true, |
| "metric_for_best_model": "eval_reward_total_mean", |
| "greater_is_better": true, |
| "hub_strategy": "checkpoint" |
| }, |
| "sanity_check": { |
| "output_root": "outputs/sanity_check", |
| "max_steps": 2, |
| "num_generations": 2, |
| "subset_size": 8, |
| "save_steps": 1, |
| "eval_steps": 1, |
| "logging_steps": 1 |
| }, |
| "logging": { |
| "save_env_snapshot": true, |
| "save_config_snapshot": true, |
| "metrics_jsonl_name": "metrics.jsonl", |
| "metrics_csv_name": "metrics.csv", |
| "train_log_name": "train.log", |
| "generations_jsonl_name": "all_generations.jsonl", |
| "checkpoint_events_name": "checkpoint_events.jsonl", |
| "split_summary_name": "split_summary.json", |
| "plots_dir_name": "plots" |
| }, |
| "watcher": { |
| "enabled": true, |
| "interval_minutes": 30, |
| "send_on_checkpoint": true, |
| "send_on_error": true |
| }, |
| "preprocess_grpo": { |
| "output_root": "outputs/preprocess_grpo", |
| "num_generations": 6, |
| "generator_batch_size": 32, |
| "max_prompt_length": 1024, |
| "max_completion_length": 640, |
| "temperature": 1.0, |
| "top_p": 1.0, |
| "progress_every": 10, |
| "vllm_gpu_memory_utilization": 0.75, |
| "vllm_tensor_parallel_size": 1, |
| "vllm_dtype": "bfloat16", |
| "vllm_max_loras": 1, |
| "vllm_max_lora_rank": 64, |
| "flashinfer_disable_version_check": true, |
| "trust_remote_code": true |
| }, |
| "preprocess_grpo_pipeline": { |
| "output_root": "outputs/preprocess_grpo_pipeline", |
| "num_generations": 6, |
| "generator_batch_size": 32, |
| "max_prompt_length": 1024, |
| "max_completion_length": 640, |
| "temperature": 1.0, |
| "top_p": 1.0, |
| "progress_every": 10, |
| "judge_poll_seconds": 20, |
| "judge_claim_stale_after_seconds": 1800, |
| "vllm_gpu_memory_utilization": 0.75, |
| "vllm_tensor_parallel_size": 1, |
| "vllm_dtype": "bfloat16", |
| "vllm_max_loras": 1, |
| "vllm_max_lora_rank": 64, |
| "flashinfer_disable_version_check": true, |
| "trust_remote_code": true |
| } |
| } |
