mr-mc/flowjudge-dialam-reconstruction-v3

FlowJudge DialAM patch reconstruction artifact

This publishable artifact documents a private educational transformation of the English DialAM/QT30 corpus for incremental argument-graph patch prediction. It contains no raw or transformed QT30 dialogue text and no original QT30 episode, map, proposition, or example IDs. Because IDs/labels-only redistribution remains unclear, identifier-bearing fields and per-example failure inventories are replaced by counts and source-file hashes. The artifact contains only allowed source/manifests, aggregate statistics, JSON schemas, frozen evaluation hashes, prompt-ceiling metrics, and deterministic reconstruction code.

The private training slices contain 256, 512, 1024, and 2048 nested examples. Every slice is an exact prefix of the next and uses an original-parent-episode split: 24 episodes for training and six untouched episodes for evaluation. The n=2048 slice contains 819 NONE, 410 SUPPORT, 307 ATTACK, 410 REPHRASE, and 102 mixed-label blocks.

All four v1 sizes and one hard-negative v2 n=2048 checkpoint were evaluated on the same frozen 30-scenario set. V1 n=2048 was the best fixed-curve point. The v2 data change added 1,024 difficult NONE blocks, but increased false edges and failed its preregistered material-improvement criterion.

The preregistered v3 experiment materially improved the behavior and is now the selected direction. Its private 4,096-row corpus contains 2,048 exact same-update positive/NONE pairs. It reserves four of the former training episodes for development, trains on the remaining 20, and keeps the original six frozen evaluation episodes untouched. Per-example assistant-token loss eliminates the output-length weighting mismatch documented after v2. V3 reached 43.3% exact patch accuracy, 35.0% edge F1, 0.300 false edges/update, and 0/6 NONE false-positive cases on the frozen set. It still does not clear the original reliability bar. The selected public adapter is `mr-mc/flowjudge-dialam-qwen3-0.6b-v3-n4096`. The text-free manifest, schema, aggregate result, and reconstruction implementation are included here; the negative reliability finding is preserved rather than presenting the model as production-ready.

Behavior

Given one new proposition and a complete fixed-size comparison block of earlier propositions from the same dialogue, return one bare JSON object containing all direct SUPPORT, ATTACK, or REPHRASE relations to supplied IDs. Return an empty list when none exists; do not emit indirect relations, invented IDs, or prose.

Reconstruction

The reconstruction/ directory contains the deterministic parser, filtering, split, prompt, freeze, and nested-corpus build code. Reconstruction requires a separately obtained official QT30 archive and permission appropriate to the user's intended use. The archive itself and generated text-bearing JSONL files are deliberately omitted.

The metadata directory records the exact source checksum, retention funnel, episode split, class counts, nested-slice hashes, frozen 30-scenario evaluation hashes, prompt-ceiling metrics, and judge-rubric fingerprint.

Redistribution boundary

Project-use approval does not establish a general public redistribution license for the underlying QT30 text. Consumers should obtain the official source and review its terms themselves. This repository is a reproducibility and evidence artifact, not a mirror of the corpus.