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

sft-flint-section-v2-flintsys arm of the caveman reasoning-compression ablation study: Qwen/Qwen3.5-4B fine-tuned (LoRA adapter) on flint/data/flint-section-aware.jsonl (551 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	0.8 (1945.4 tok, loops 0.09)	0.575 (4413.4 tok, loops 0.2167)
math500@t0.0	0.59 (4384.2 tok, loops 0.19)	0.3083 (6631.6 tok, loops 0.2833)

Run manifest

{
  "arm": "sft-flint-section-v2-flintsys",
  "dataset": "flint/data/flint-section-aware.jsonl",
  "rows": 551,
  "dropped_overlong": 1,
  "epochs": 2,
  "system_prompts": true,
  "system_file": "ember/configs/system-flint.txt",
  "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": 5658.2296,
  "final_loss": 0.2527087450027466,
  "log_history": [
    {
      "loss": 0.3424527168273926,
      "grad_norm": 0.1815948486328125,
      "learning_rate": 0.000196068518757684,
      "epoch": 0.29038112522686027,
      "step": 10
    },
    {
      "loss": 0.3740869998931885,
      "grad_norm": 0.16519610583782196,
      "learning_rate": 0.0001731531335263669,
      "epoch": 0.5807622504537205,
      "step": 20
    },
    {
      "loss": 0.31851556301116946,
      "grad_norm": 0.1333686113357544,
      "learning_rate": 0.0001344466850284333,
      "epoch": 0.8711433756805808,
      "step": 30
    },
    {
      "loss": 0.2851787328720093,
      "grad_norm": 0.1366184800863266,
      "learning_rate": 8.830446780279176e-05,
      "epoch": 1.1451905626134302,
      "step": 40
    },
    {
      "loss": 0.23676638603210448,
      "grad_norm": 0.12336449325084686,
      "learning_rate": 4.468688458748006e-05,
      "epoch": 1.4355716878402904,
      "step": 50
    },
    {
      "loss": 0.23735339641571046,
      "grad_norm": 0.13725954294204712,
      "learning_rate": 1.300936275912098e-05,
      "epoch": 1.7259528130671506,
      "step": 60
    },
    {
      "loss": 0.2527087450027466,
      "grad_norm": 0.26777997612953186,
      "learning_rate": 1.0991085142886271e-07,
      "epoch": 2.0,
      "step": 70
    },
    {
      "train_runtime": 5658.2296,
      "train_samples_per_second": 0.195,
      "train_steps_per_second": 0.012,
      "total_flos": 1.5344224468149658e+17,
      "train_loss": 0.2924375057220459,
      "epoch": 2.0,
      "step": 70
    }
  ]
}