| { |
| "model_name": "Grug-12B", |
| "base_model": "google/gemma-4-12B-it", |
| "artifact_type": "merged_transformers_safetensors_model", |
| "peft_adapter_merged": true, |
| "teacher_model_for_compact_reasoning": "cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit", |
| "teacher_runtime": "vllm", |
| "dataset_counts": { |
| "normalized_recent_reasoning": 8680, |
| "selected_verbose": 6144, |
| "compact_raw": 6144, |
| "compact_verified": 5740, |
| "compact_rejected": 404, |
| "packed_train": 5166, |
| "packed_validation": 287, |
| "packed_test": 287 |
| }, |
| "verified_source_counts": { |
| "hotdogs/uka-glm-5.2": 1617, |
| "Scale-or-Reason/general-reasoning-ift-pairs": 1305, |
| "samcheng0/lumia-reasoning-sft-v1": 1103, |
| "HSH-Intelligence/verified-math-reasoning-3k": 672, |
| "kd13/CodeDebug-Instruct-v2-Reasoning": 600, |
| "Madarabr/cortex-adaptive-thinking": 300, |
| "CL-From-Nothing/code_rose_initial_1_7B_SFT_10K_rollouts_Qwen3-4B-Thinking-2507_k12_t0.7_maxtok12288": 143 |
| }, |
| "training": { |
| "method": "QLoRA completion-only SFT", |
| "quantization": "4-bit NF4, BF16 compute", |
| "max_seq_length": 6144, |
| "max_steps": 100, |
| "batch_size": 1, |
| "gradient_accumulation_steps": 16, |
| "learning_rate": 0.00008, |
| "lora_r": 16, |
| "lora_alpha": 32, |
| "lora_dropout": 0.05, |
| "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], |
| "final_eval_loss": 0.8895471096038818, |
| "train_runtime_seconds": 2121 |
| }, |
| "local_eval": { |
| "name": "EOS-only math proxy eval", |
| "generation_mode": "eos_only", |
| "max_new_tokens": null, |
| "timeout_seconds": 900, |
| "rows": 36, |
| "base": { |
| "total_generated_tokens": 8227, |
| "avg_generated_tokens": 228.52777777777777, |
| "proxy_accuracy": 0.9166666666666666, |
| "contains_expected_rate": 0.9166666666666666, |
| "numeric_last_match_rate": 0.8611111111111112 |
| }, |
| "adapter": { |
| "total_generated_tokens": 2482, |
| "avg_generated_tokens": 68.94444444444444, |
| "proxy_accuracy": 1.0, |
| "contains_expected_rate": 1.0, |
| "numeric_last_match_rate": 1.0 |
| } |
| }, |
| "acknowledgement": "Thanks to Lambda (https://lambda.com/), the inference provider, for compute credits." |
| } |
|
|
