qwen3.5-4b-sft-flint-mixed-v2

qwen3.5-4b-sft-flint-mixed-v2

sft-flint-mixed-v2 arm of the caveman reasoning-compression ablation study: Qwen/Qwen3.5-4B fine-tuned (LoRA adapter) on flint/data/flint-mixed.jsonl (571 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-flint-mixed-v2",
  "dataset": "flint/data/flint-mixed.jsonl",
  "rows": 571,
  "dropped_overlong": 0,
  "epochs": 2,
  "system_prompts": true,
  "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": 4257.3918,
  "final_loss": 0.20572218894958497,
  "log_history": [
    {
      "loss": 0.3235448122024536,
      "grad_norm": 0.1792268007993698,
      "learning_rate": 0.00019629172873477995,
      "epoch": 0.28021015761821366,
      "step": 10
    },
    {
      "loss": 0.30161023139953613,
      "grad_norm": 0.1683519035577774,
      "learning_rate": 0.00017461836858476856,
      "epoch": 0.5604203152364273,
      "step": 20
    },
    {
      "loss": 0.2957916736602783,
      "grad_norm": 0.12694813311100006,
      "learning_rate": 0.00013774191957526143,
      "epoch": 0.840630472854641,
      "step": 30
    },
    {
      "loss": 0.27068142890930175,
      "grad_norm": 0.16031979024410248,
      "learning_rate": 9.317575866353292e-05,
      "epoch": 1.1120840630472855,
      "step": 40
    },
    {
      "loss": 0.2206240177154541,
      "grad_norm": 0.14184938371181488,
      "learning_rate": 5.000000000000002e-05,
      "epoch": 1.3922942206654991,
      "step": 50
    },
    {
      "loss": 0.2086726188659668,
      "grad_norm": 0.17701619863510132,
      "learning_rate": 1.7011470800097496e-05,
      "epoch": 1.6725043782837128,
      "step": 60
    },
    {
      "loss": 0.20572218894958497,
      "grad_norm": 0.15691602230072021,
      "learning_rate": 9.314053963669245e-07,
      "epoch": 1.9527145359019265,
      "step": 70
    },
    {
      "train_runtime": 4257.3918,
      "train_samples_per_second": 0.268,
      "train_steps_per_second": 0.017,
      "total_flos": 1.520973357373317e+17,
      "train_loss": 0.25957077369093895,
      "epoch": 2.0,
      "step": 72
    }
  ]
}