flint-mixed-qwen3.5-4b

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

Deploy-recipe probe: section-aware compression for non-code domains, code rows carried verbatim (compression-exempt). Built by build_mixed.py from the section-aware variant + raw crucible code rows. Outcome: falsified — mixing verbatim code rows next to compressed math AMPLIFIES code-thinking length at eval.

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: 571 — domains: {"general": 192, "math": 161, "creative": 117, "code": 101}
• Mean compression ratio (caveman/raw tokens): 0.68
• 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": "2026-07-09T20:40:32+00:00",
  "git_sha": "11533e2cdea835aaf8a36354eed9fadd3ce592de",
  "variant": "mixed",
  "policy": "section-aware for non-code (from flint-section-aware.jsonl), verbatim crucible traces for code",
  "section_input": "/mnt/data/projects/reasoning-compression/flint/data/flint-section-aware.jsonl",
  "crucible_input": "/mnt/data/projects/reasoning-compression/crucible/data/crucible-sft-qwen3.5-4b-mini.jsonl",
  "rows": 571,
  "domains": {
    "general": 192,
    "math": 161,
    "creative": 117,
    "code": 101
  }
}