Upload configs/rae_training_config.json with huggingface_hub
Browse files
configs/rae_training_config.json
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": {
|
| 3 |
+
"base_model": "Qwen/Qwen2.5-7B-Instruct",
|
| 4 |
+
"quantization": "int4",
|
| 5 |
+
"torch_dtype": "bfloat16",
|
| 6 |
+
"attn_implementation": "flash_attention_2",
|
| 7 |
+
"trust_remote_code": true
|
| 8 |
+
},
|
| 9 |
+
"lora": {
|
| 10 |
+
"enabled": true,
|
| 11 |
+
"r": 32,
|
| 12 |
+
"alpha": 64,
|
| 13 |
+
"dropout": 0.05,
|
| 14 |
+
"target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
|
| 15 |
+
"task_type": "CAUSAL_LM",
|
| 16 |
+
"bias": "none"
|
| 17 |
+
},
|
| 18 |
+
"data": {
|
| 19 |
+
"train_path": "data/rae_training_data/train.jsonl",
|
| 20 |
+
"eval_path": "data/rae_training_data/validation.jsonl",
|
| 21 |
+
"max_seq_length": 4096,
|
| 22 |
+
"chat_template": "auto",
|
| 23 |
+
"phase_tags": {
|
| 24 |
+
"saturation": ["<SATURATION>", "</SATURATION>"],
|
| 25 |
+
"abstraction": ["<ABSTRACTION>", "</ABSTRACTION>"],
|
| 26 |
+
"descent": ["<DESCENT>", "</DESCENT>"],
|
| 27 |
+
"integration": ["<INTEGRATION>", "</INTEGRATION>"]
|
| 28 |
+
}
|
| 29 |
+
},
|
| 30 |
+
"training": {
|
| 31 |
+
"epochs": 3,
|
| 32 |
+
"batch_size": 1,
|
| 33 |
+
"gradient_accumulation_steps": 8,
|
| 34 |
+
"effective_batch_size": 8,
|
| 35 |
+
"learning_rate": 5e-6,
|
| 36 |
+
"lr_scheduler": "cosine",
|
| 37 |
+
"warmup_ratio": 0.1,
|
| 38 |
+
"weight_decay": 0.01,
|
| 39 |
+
"max_grad_norm": 1.0,
|
| 40 |
+
"bf16": true,
|
| 41 |
+
"logging_steps": 10,
|
| 42 |
+
"eval_steps": 100,
|
| 43 |
+
"save_steps": 200,
|
| 44 |
+
"save_total_limit": 3
|
| 45 |
+
},
|
| 46 |
+
"rae_loss": {
|
| 47 |
+
"_comment": "Multi-objective loss replicating the handwriting effect",
|
| 48 |
+
"enabled": true,
|
| 49 |
+
"phase_weights": {
|
| 50 |
+
"saturation": 1.0,
|
| 51 |
+
"abstraction": 1.5,
|
| 52 |
+
"descent": 1.5,
|
| 53 |
+
"integration": 1.0
|
| 54 |
+
},
|
| 55 |
+
"coherence_weight": 0.3,
|
| 56 |
+
"compression_weight": 0.2,
|
| 57 |
+
"_coherence_note": "Penalizes abstraction that doesn't follow from saturation",
|
| 58 |
+
"_compression_note": "Rewards abstraction being shorter than saturation (information compression)"
|
| 59 |
+
},
|
| 60 |
+
"output": {
|
| 61 |
+
"dir": "outputs/rae-trained-model",
|
| 62 |
+
"push_to_hub": true,
|
| 63 |
+
"hub_model_id": "rae-cognitive-model",
|
| 64 |
+
"hub_token_env": "HF_TOKEN"
|
| 65 |
+
},
|
| 66 |
+
"wandb": {
|
| 67 |
+
"enabled": true,
|
| 68 |
+
"project": "rae-training",
|
| 69 |
+
"name": "rae-handwriting-v1"
|
| 70 |
+
}
|
| 71 |
+
}
|