{ "base_model": "LiquidAI/LFM2.5-2.6B", "train_path": "/home/jason/.tyness_train/datasets/boe-focused-adv-v4/train.jsonl", "output_dir": "/home/jason/.tyness_train/runs/lfm25-26b-ts-adv-5ep", "hardware": { "compute_dtype": "float32", "profile": "gtx1060-6gb" }, "preset": { "description": "Default QLoRA SFT preset for 1B-2B models on GTX 1060 6GB. Good quality/speed balance.", "method": "sft-qlora", "target_model_size": "1B-2B", "quantization": "nf4", "compute_dtype": "float32", "lora_r": 16, "lora_alpha": 32, "lora_dropout": 0.05, "target_modules": [ "q_proj", "k_proj", "v_proj", "o_proj" ], "learning_rate": 0.0002, "epochs": 5, "batch_size": 1, "grad_accum": 16, "seq_len": 2048, "warmup_ratio": 0.03, "scheduler": "cosine", "gradient_checkpointing": 1, "min_vram_gb": 6.0, "notes": "Effective batch 16, seq 2304 fp32 ceiling (run19-proven recipe, restored 2026-10-12). Liquid AI recommends r=16, alpha=2r for LFM2/2.5 SFT.", "prepend_empty_think": true } }