Fine-tuned Llama 3.1 8B adapter for structured API generation - 40% accuracy (95% better than Azure GPT)
02e3cea
verified
| { | |
| "model_id": "unsloth/llama-3.1-8b-instruct-bnb-4bit", | |
| "approach": "llama_cep_pure_peft", | |
| "cep_type": "universal", | |
| "timestamp": "2025-11-12T14:30:08.645407", | |
| "training_duration": "0:04:52.278900", | |
| "hyperparams": { | |
| "lora_r": 32, | |
| "lora_alpha": 64, | |
| "lora_dropout": 0.05, | |
| "batch_size": 2, | |
| "grad_accum_steps": 4, | |
| "learning_rate": 0.0002, | |
| "epochs": 3, | |
| "max_seq_length": 512 | |
| }, | |
| "dataset": { | |
| "train_path": "slm_swap/02_dataset/structured/train_chat.jsonl", | |
| "train_size": 300, | |
| "val_path": "slm_swap/02_dataset/structured/val_chat.jsonl", | |
| "val_size": 60 | |
| }, | |
| "trainer_stats": "TrainOutput(global_step=114, training_loss=0.5034574193128368, metrics={'train_runtime': 291.9495, 'train_samples_per_second': 3.083, 'train_steps_per_second': 0.39, 'total_flos': 6487886798585856.0, 'train_loss': 0.5034574193128368, 'epoch': 3.0})", | |
| "cep": "<|formatting_rules|>\nCRITICAL OUTPUT RULES - FOLLOW EXACTLY:\n\n1. JSON Outputs:\n - Output ONLY raw JSON object\n - NO markdown blocks (no ```json or ```)\n - NO explanatory text before or after\n - Match parameter names EXACTLY as in signature\n - Use proper JSON types: strings \"...\", numbers 123, arrays [...]\n\n2. XML Tool Calls:\n - Format: <tool_call name=\"FUNC_NAME\">{\"arg\": \"value\"}</tool_call>\n - MUST have opening tag: <tool_call name=\"...\">\n - MUST have closing tag: </tool_call>\n - NO self-closing tags (NOT <tool_call ... />)\n - NO \"arguments=\" attribute\n - JSON dict inside tags\n\n3. Common Errors to AVOID:\n \u274c ```json{\"key\": \"value\"}``` (markdown wrapper)\n \u274c <tool_call arguments='...'> (wrong attribute)\n \u274c <tool_call name=\"func\" ... /> (self-closing)\n \u274c {\"query\": \"x\"} when signature says \"q\" (param mismatch)\n \u274c {\"country\": [\"US\",\"CA\"]} when expecting \"US,CA\" (array vs string)\n\n4. Parameter Mapping:\n - \"q\" NOT \"query\"\n - \"is_id\" NOT \"id\"\n - Check signature for EXACT names\n - Include ALL required fields\n - Omit optional fields if not needed\n\nVERIFY: Does output match format rules? Check before responding.\n<|end_formatting_rules|>\n\n", | |
| "note": "Pure Transformers + PEFT (bypassed Unsloth compatibility issue)" | |
| } |