flint-flat-adaptive-qwen3.5-4b

Compressed ("caveman") reasoning traces for SFT — the flat-adaptive variant of the flint reasoning-compression pipeline. Converted from verified self-distilled traces by Qwen/Qwen3.5-4B, policy policy/1.1, template caveman_convert/2.0.

Flat compression, level adaptive to trace length (short = heavy, long = light), wind-down tail preserved verbatim (policy 1.1).

Each row: input, reasoning (compressed trace), answer (carried verbatim from the source), domain, verified, difficulty, and meta with per-row provenance (converter/segmenter model, span labels + actions for section-aware rows, token counts raw vs compressed).

• Rows: 332 — domains: {"general": 110, "math": 106, "creative": 105, "code": 11}
• Mean compression ratio (caveman/raw tokens): 0.35
• Content-preservation linters (numbers, negations, byte-identical code blocks, repetition) reject unsafe conversions; rejects are excluded.

Part of the caveman reasoning-compression ablation study.

Build manifest

{
  "created_at": "2026-07-09T12:09:04+00:00",
  "git_sha": "cb1eaae9ddc9ba9acd8c5554488d2fdf1d3cf52e",
  "variant": "flat-adaptive",
  "config": {
    "mode": "flat",
    "level": "adaptive"
  },
  "counts": {
    "ok": 332,
    "rejected": 316,
    "failed": 0
  },
  "convert_model": "Qwen/Qwen3.5-4B",
  "segment_model": null,
  "template_version": "caveman_convert/2.0",
  "policy_version": "policy/1.1",
  "segment_template_version": null,
  "tokenizer": "Qwen/Qwen3.5-4B",
  "input_rows": 648
}