Modelpublic

marcodsn/qwen3.5-4b-sft-flint-section37-v2

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes

Model Card

qwen3.5-4b-sft-flint-section37-v2

sft-flint-section37-v2 arm of the caveman reasoning-compression ablation study: Qwen/Qwen3.5-4B fine-tuned (LoRA adapter) on flint/data/flint-section-aware-step37.jsonl (638 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.

suite	this model	original Qwen3.5-4B
gsm8k@t0.0	0.59 (3934.9 tok, loops 0.25)	0.575 (4413.4 tok, loops 0.2167)
humaneval@t0.0	0.59 (4989.1 tok, loops 0.17)	0.9083 (821.5 tok, loops 0.0083)
loops:gsm8k@t0.0	0.54 (4198.7 tok, loops 0.36)	0.64 (3915.1 tok, loops 0.28)
loops:gsm8k@t0.6	0.68 (3509.3 tok, loops 0.16)	0.68 (3827.4 tok, loops 0.12)
loops:gsm8k@t1.0	0.76 (3989.7 tok, loops 0.1)	0.74 (4140.6 tok, loops 0.0)
math500@t0.0	0.3 (6560.5 tok, loops 0.31)	0.3083 (6631.6 tok, loops 0.2833)
mbpp@t0.0	0.48 (6213.1 tok, loops 0.22)	0.75 (1564.8 tok, loops 0.0083)

Run manifest

json

{
  "arm": "sft-flint-section37-v2",
  "dataset": "flint/data/flint-section-aware-step37.jsonl",
  "rows": 638,
  "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": 5232.8667,
  "final_loss": 0.26767816543579104,
  "log_history": [
    {
      "loss": 0.3759377717971802,
      "grad_norm": 0.20056962966918945,
      "learning_rate": 0.00019701859555740648,
      "epoch": 0.2507836990595611,
      "step": 10
    },
    {
      "loss": 0.3634188175201416,
      "grad_norm": 0.14793893694877625,
      "learning_rate": 0.00017943870854121124,
      "epoch": 0.5015673981191222,
      "step": 20
    },
    {
      "loss": 0.35701828002929686,
      "grad_norm": 0.1765201836824417,
      "learning_rate": 0.00014881762060482814,
      "epoch": 0.7523510971786834,
      "step": 30
    },
    {
      "loss": 0.34912691116333006,
      "grad_norm": 0.1738613247871399,
      "learning_rate": 0.00011018229867038356,
      "epoch": 1.0,
      "step": 40
    },
    {
      "loss": 0.26037137508392333,
      "grad_norm": 0.1503107249736786,
      "learning_rate": 6.98753810462766e-05,
      "epoch": 1.250783699059561,
      "step": 50
    },
    {
      "loss": 0.2628163814544678,
      "grad_norm": 0.1447693109512329,
      "learning_rate": 3.45139266054715e-05,
      "epoch": 1.5015673981191222,
      "step": 60
    },
    {
      "loss": 0.26981985569000244,
      "grad_norm": 0.15801295638084412,
      "learning_rate": 9.903113209758096e-06,
      "epoch": 1.7523510971786833,
      "step": 70
    },
    {
      "loss": 0.26767816543579104,
      "grad_norm": 0.18598252534866333,
      "learning_rate": 8.322006554171146e-08,
      "epoch": 2.0,
      "step": 80
    },
    {
      "train_runtime": 5232.8667,
      "train_samples_per_second": 0.244,
      "train_steps_per_second": 0.015,
      "total_flos": 1.872236310679388e+17,
      "train_loss": 0.31327344477176666,
      "epoch": 2.0,
      "step": 80
    }
  ]
}