qwen3.5-4b-sft-crucible-v2

qwen3.5-4b-sft-crucible-v2

sft-crucible-v2 arm of the caveman reasoning-compression ablation study: Qwen/Qwen3.5-4B fine-tuned (LoRA adapter) on crucible/data/crucible-sft-qwen3.5-4b-mini.jsonl (646 rows, 2 epochs, LoRA r=64).

The study asks whether compressed ("caveman") reasoning traces can train a model to reason in fewer tokens without losing accuracy — and which parts of a trace are compressible. See the run manifest below for the exact recipe; eval results live in the study's report.

Eval summary (t=0, max_tokens 8192)

Accuracy (avg reasoning tokens, loop rate) — this arm vs the original model it was fine-tuned from (Qwen/Qwen3.5-4B), same harness and prompts.

Run manifest

{
  "arm": "sft-crucible-v2",
  "dataset": "crucible/data/crucible-sft-qwen3.5-4b-mini.jsonl",
  "rows": 646,
  "dropped_overlong": 2,
  "epochs": 2,
  "lora": {
    "r": 64,
    "alpha": 128,
    "dropout": 0.0,
    "target": "all"
  },
  "train": {
    "epochs": 2,
    "lr": 0.0002,
    "batch_size": 1,
    "grad_accum": 16,
    "warmup_ratio": 0.03,
    "lr_scheduler": "cosine",
    "weight_decay": 0.01,
    "seed": 3407,
    "logging_steps": 10,
    "save_strategy": "epoch"
  },
  "model": {
    "name": "Qwen/Qwen3.5-4B",
    "max_seq_length": 13312,
    "load_in_4bit": true,
    "chat_template": "qwen3.5"
  },
  "train_runtime_s": 8177.9355,
  "final_loss": 0.15754287242889403,
  "log_history": [
    {
      "loss": 0.26338093280792235,
      "grad_norm": 0.18449434638023376,
      "learning_rate": 0.000197166934004041,
      "epoch": 0.2476780185758514,
      "step": 10
    },
    {
      "loss": 0.21944773197174072,
      "grad_norm": 0.1204717829823494,
      "learning_rate": 0.00018043165652707649,
      "epoch": 0.4953560371517028,
      "step": 20
    },
    {
      "loss": 0.2305692195892334,
      "grad_norm": 0.144414484500885,
      "learning_rate": 0.00015114354791034225,
      "epoch": 0.7430340557275542,
      "step": 30
    },
    {
      "loss": 0.22316784858703614,
      "grad_norm": 0.11685270071029663,
      "learning_rate": 0.00011387355319890685,
      "epoch": 0.9907120743034056,
      "step": 40
    },
    {
      "loss": 0.16748157739639283,
      "grad_norm": 0.11084543913602829,
      "learning_rate": 7.443833675595255e-05,
      "epoch": 1.2229102167182662,
      "step": 50
    },
    {
      "loss": 0.16777762174606323,
      "grad_norm": 0.12330946326255798,
      "learning_rate": 3.899248539894757e-05,
      "epoch": 1.4705882352941178,
      "step": 60
    },
    {
      "loss": 0.16063579320907592,
      "grad_norm": 0.13280856609344482,
      "learning_rate": 1.3067972556041752e-05,
      "epoch": 1.718266253869969,
      "step": 70
    },
    {
      "loss": 0.15754287242889403,
      "grad_norm": 0.12890632450580597,
      "learning_rate": 7.10792629802659e-07,
      "epoch": 1.9659442724458205,
      "step": 80
    },
    {
      "train_runtime": 8177.9355,
      "train_samples_per_second": 0.158,
      "train_steps_per_second": 0.01,
      "total_flos": 2.1623255705689498e+17,
      "train_loss": 0.19756304926988555,
      "epoch": 2.0,
      "step": 82
    }
  ]
}