mr-mc/flowjudge-dialam

FlowJudge DialAM incremental argument patches

This is the actual transformed dataset used to test whether a 0.6B open model can learn one falsifiable behavior: given one new proposition and one complete fixed-size block of earlier propositions from the same dialogue, emit every and only direct SUPPORT, ATTACK, or REPHRASE edge as one bare JSON object. An empty relation list is required when no direct edge exists.

The package includes the nested v1 data-efficiency curve, the v2 hard-negative intervention, the v3 paired patch corpus, the selected v5/v5.1 pairwise pipeline, the project-owned development and frozen evaluation sets, schemas, aggregate reports, a chronological improvement chart, and deterministic reconstruction code. The evidence/ directory preserves every prompt-ceiling candidate/judge record plus raw candidate and blinded-judge JSONL for the base, all four v1 checkpoints, v2, v3, and the promoted v5.1 result.

Splits and leakage controls

All splits operate on original QT30 parent episodes, never individual DialAM map IDs. V1/v2 use 24 training episodes and six frozen evaluation episodes. V3/v5 reserve four of those former training episodes as a separate development set, train on the remaining 20, and leave the same six frozen episodes untouched. The 30-case development and 30-case frozen files are project-owned evaluation scenarios; the grader's staff-held-out set is not included.

The v1 sizes are deterministic nested prefixes. V3 has 4,096 block-level rows arranged as 2,048 exact same-update positive/NONE contrast pairs. V5 has 8,192 candidate-level rows arranged as 4,096 positive/NONE pairs: 1,366 ATTACK, 1,365 REPHRASE, 1,365 SUPPORT, and 4,096 NONE rows. V5 still supplies the complete comparison block in every input; its fixed four-label predictions are assembled deterministically into the external block-level JSON contract. V5.1 keeps the same checkpoint and applies the preregistered fixed 3.0 NONE margin selected on the separate development set.

Source and permission

The source is English QT30 as distributed for DialAM-2024 by ARG-tech at the University of Dundee. Exact archive URL, retrieval metadata, SHA-256 hashes, format references, and known archive/API discrepancies are in metadata/source_manifest.json and metadata/DATA_AUDIT.md.

The project owner directly attested on 2026-08-23 that their project-specific permission includes redistribution. This repository therefore publishes the curated transformed training/evaluation JSONL and original identifiers needed for provenance. It does not claim a general QT30 license or grant rights beyond this project. The official raw archive and extracted maps are deliberately not mirrored; obtain those from ARG-tech.

Reproduction

From the FlowJudge source repository, place the separately obtained official archive at the documented input path and run:

uv sync --frozen
uv run python scripts/build_dialam_gate.py
uv run python scripts/validate_dialam_gate.py
uv run python scripts/build_dialam_training.py
uv run python scripts/build_dialam_training_v2.py
uv run python scripts/build_dialam_training_v3.py
uv run python scripts/build_dialam_training_v5.py
uv run python scripts/prepare_dialam_hf_dataset.py

Every published file is inventoried with its byte count and SHA-256 hash in publish_manifest.json. The manifests record the seed, exact episode split, class mix, filter funnel, frozen eval/rubric hashes, training configuration, checkpoint evidence, and all preregistered experiment decisions.

Limitations

The selected v5.1 result reaches 53.3% exact patches, 47.6% edge F1, and 47.4% relation macro-F1 on the reused 30-scenario frozen benchmark. It passes the project's preregistered v3-to-v5.1 promotion gate but not the original high reliability bar; it is not production-ready.

The relation labels inherit QT30's annotations and the documented grounding, chronology, and binary-relation filters. The frozen set has only 30 scenarios, so small absolute changes produce visibly large percentage changes. Failed experiments remain in the history to prevent selective reporting. Consult the latest result and promotion decision in metadata/dialam_v5_1_results.md; do not interpret publication as a claim that the selected model is production-ready.