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

sft-flint-section-v2 arm of the caveman reasoning-compression ablation study: Qwen/Qwen3.5-4B fine-tuned (LoRA adapter) on flint/data/flint-section-aware.jsonl (552 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
creative@t0.0	None (3609.3 tok, loops 0.23)	None (3958.4 tok, loops 0.26)
gsm8k@t0.0	0.77 (2585.7 tok, loops 0.11)	0.575 (4413.4 tok, loops 0.2167)
humaneval@t0.0	0.53 (5046.8 tok, loops 0.16)	0.9083 (821.5 tok, loops 0.0083)
loops:gsm8k@t0.0	0.82 (2210.6 tok, loops 0.14)	0.64 (3915.1 tok, loops 0.28)
loops:gsm8k@t0.6	0.84 (2266.3 tok, loops 0.08)	0.68 (3827.4 tok, loops 0.12)
loops:gsm8k@t1.0	0.78 (2609.5 tok, loops 0.08)	0.74 (4140.6 tok, loops 0.0)
math500@t0.0	0.53 (5120.1 tok, loops 0.16)	0.3083 (6631.6 tok, loops 0.2833)
mbpp@t0.0	0.41 (6090.1 tok, loops 0.32)	0.75 (1564.8 tok, loops 0.0083)
mmlu_pro@t0.0	0.52 (4426.6 tok, loops 0.14)	0.6083 (4654.4 tok, loops 0.0583)

Run manifest

{
  "arm": "sft-flint-section-v2",
  "dataset": "flint/data/flint-section-aware.jsonl",
  "rows": 552,
  "dropped_overlong": 0,
  "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": 3658.9651,
  "final_loss": 0.24562442302703857,
  "log_history": [
    {
      "loss": 0.3610557794570923,
      "grad_norm": 0.22741258144378662,
      "learning_rate": 0.000196068518757684,
      "epoch": 0.2898550724637681,
      "step": 10
    },
    {
      "loss": 0.3423754215240479,
      "grad_norm": 0.18591611087322235,
      "learning_rate": 0.0001731531335263669,
      "epoch": 0.5797101449275363,
      "step": 20
    },
    {
      "loss": 0.3451270341873169,
      "grad_norm": 0.17014743387699127,
      "learning_rate": 0.0001344466850284333,
      "epoch": 0.8695652173913043,
      "step": 30
    },
    {
      "loss": 0.26852803230285643,
      "grad_norm": 0.13987186551094055,
      "learning_rate": 8.830446780279176e-05,
      "epoch": 1.144927536231884,
      "step": 40
    },
    {
      "loss": 0.23699781894683838,
      "grad_norm": 0.16187669336795807,
      "learning_rate": 4.468688458748006e-05,
      "epoch": 1.434782608695652,
      "step": 50
    },
    {
      "loss": 0.2425454378128052,
      "grad_norm": 0.1568235158920288,
      "learning_rate": 1.300936275912098e-05,
      "epoch": 1.7246376811594204,
      "step": 60
    },
    {
      "loss": 0.24562442302703857,
      "grad_norm": 0.2882634997367859,
      "learning_rate": 1.0991085142886271e-07,
      "epoch": 2.0,
      "step": 70
    },
    {
      "train_runtime": 3658.9651,
      "train_samples_per_second": 0.302,
      "train_steps_per_second": 0.019,
      "total_flos": 1.3342011412702003e+17,
      "train_loss": 0.2917505638939994,
      "epoch": 2.0,
      "step": 70
    }
  ]
}