flint-mixed-qwen3.5-4b

# 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 ```json { "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 } } ```