Instructions to use brikdavies/qwen1.7B-MMLU-hint-following-RL with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use brikdavies/qwen1.7B-MMLU-hint-following-RL with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("brikdavies/qwen1.7B-MMLU-hint-following-RL", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Upload nohint_config.json with huggingface_hub
Browse files- nohint_config.json +73 -0
nohint_config.json
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_name": "Qwen/Qwen3-1.7B",
|
| 3 |
+
"init_checkpoint": "",
|
| 4 |
+
"ref_model_name": "",
|
| 5 |
+
"skip_first_n_prompts": 0,
|
| 6 |
+
"dataset_name": "cais/mmlu",
|
| 7 |
+
"mmlu_config": "all",
|
| 8 |
+
"mmlu_split": "test",
|
| 9 |
+
"lambda_length": 0.0,
|
| 10 |
+
"format_penalty": -1.0,
|
| 11 |
+
"max_steps": 250,
|
| 12 |
+
"per_device_train_batch_size": 0,
|
| 13 |
+
"gradient_accumulation_steps": 6,
|
| 14 |
+
"learning_rate": 5e-05,
|
| 15 |
+
"warmup_ratio": 0.05,
|
| 16 |
+
"lr_scheduler_type": "constant_with_warmup",
|
| 17 |
+
"max_grad_norm": 1.0,
|
| 18 |
+
"bf16": true,
|
| 19 |
+
"gradient_checkpointing": true,
|
| 20 |
+
"num_generations": 6,
|
| 21 |
+
"max_completion_length": 3000,
|
| 22 |
+
"max_prompt_length": 512,
|
| 23 |
+
"temperature": 0.6,
|
| 24 |
+
"top_p": 0.95,
|
| 25 |
+
"loss_type": "dr_grpo",
|
| 26 |
+
"beta": 0.0,
|
| 27 |
+
"num_iterations": 1,
|
| 28 |
+
"epsilon": 0.2,
|
| 29 |
+
"scale_rewards": "group",
|
| 30 |
+
"lora_rank": 16,
|
| 31 |
+
"lora_alpha": 32,
|
| 32 |
+
"lora_dropout": 0.05,
|
| 33 |
+
"lora_target_modules": [
|
| 34 |
+
"q_proj",
|
| 35 |
+
"k_proj",
|
| 36 |
+
"v_proj",
|
| 37 |
+
"o_proj",
|
| 38 |
+
"gate_proj",
|
| 39 |
+
"up_proj",
|
| 40 |
+
"down_proj"
|
| 41 |
+
],
|
| 42 |
+
"use_vllm": true,
|
| 43 |
+
"vllm_mode": "colocate",
|
| 44 |
+
"vllm_gpu_memory_utilization": 0.2,
|
| 45 |
+
"vllm_model_impl": "transformers",
|
| 46 |
+
"logging_steps": 1,
|
| 47 |
+
"log_completions": false,
|
| 48 |
+
"report_to": "wandb",
|
| 49 |
+
"save_steps": 50,
|
| 50 |
+
"save_total_limit": 5,
|
| 51 |
+
"output_dir": "./outputs",
|
| 52 |
+
"run_name": "mmlu_nohint_correctness_lr5e5_250steps",
|
| 53 |
+
"seed": 42,
|
| 54 |
+
"baseline_num_problems": 100,
|
| 55 |
+
"baseline_generations_per_problem": 8,
|
| 56 |
+
"wandb_project": "grpo-mmlu-nohint",
|
| 57 |
+
"lora_config": {
|
| 58 |
+
"r": 16,
|
| 59 |
+
"lora_alpha": 32,
|
| 60 |
+
"lora_dropout": 0.05,
|
| 61 |
+
"target_modules": [
|
| 62 |
+
"gate_proj",
|
| 63 |
+
"k_proj",
|
| 64 |
+
"down_proj",
|
| 65 |
+
"q_proj",
|
| 66 |
+
"up_proj",
|
| 67 |
+
"o_proj",
|
| 68 |
+
"v_proj"
|
| 69 |
+
],
|
| 70 |
+
"bias": "none",
|
| 71 |
+
"task_type": "TaskType.CAUSAL_LM"
|
| 72 |
+
}
|
| 73 |
+
}
|