Modelpublic

mr-mc/flowjudge-dialam-qwen3-0.6b-v5-1-n8192

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes11downloads

Model Card

Qwen3-0.6B DialAM pairwise patch adapter v5.1 / n=8192

This is the selected FlowJudge QLoRA adapter for one narrow behavior: given one new proposition and a complete block of earlier propositions, emit every and only direct SUPPORT, ATTACK, or REPHRASE edge to supplied IDs as one bare JSON object. Return an empty relation list when no direct edge exists.

V5 trains 8,192 candidate-level rows as 4,096 exact positive/NONE contrast pairs. V5.1 does not retrain or alter the checkpoint. It applies the preregistered fixed 3.0 NONE margin selected on a separate four-parent-episode development set to correct the known 50.0% training versus 7.379% natural positive-prior shift.

Required inference contract

This adapter is a fixed-label classifier, not an unconstrained text generator. For every earlier candidate, render the complete block with the v5 pairwise prompt, compute the mean assistant-token log probability of NONE, SUPPORT, ATTACK, and REPHRASE, and apply dialam_inference_config.json. Emitting the result of ordinary .generate() is not equivalent. The source repository's eval.py implements the exact pipeline and deterministic block assembly.

bash

uv sync --group train
uv run python eval.py \
  --model mr-mc/flowjudge-dialam-qwen3-0.6b-v5-1-n8192 \
  --eval-set <dialam-patch-example-jsonl>

Frozen 30-scenario result

Metric	Previous v3 / 4096	Selected v5.1 / 8192
Exact patch accuracy	43.3%	53.3%
Edge precision	43.8%	55.6%
Edge recall	29.2%	41.7%
Edge F1	35.0%	47.6%
Relation macro-F1	33.9%	47.4%
ATTACK F1	18.2%	30.8%
False edges/update	0.300	0.267
NONE cases with false edges	0/6	0/6
Judge Robustness /4	2.967	3.033

V5.1 passed all nine preregistered reused-benchmark promotion checks and is the selected submission model. It still fails the original production-like reliability bar (80% exact, 85% edge F1, 75% macro-F1, at most 0.20 false edges/update, and 3.5/4 Robustness), so this is an educational research artifact, not a production-ready argument-mining claim.

Source and permission

Training data is transformed from English QT30 as distributed for DialAM-2024. The project owner directly attested project-specific use and redistribution permission on 2026-08-23. No general QT30 license is claimed. The official raw archive/maps are not included; the companion dataset repository publishes the permission-cleared transformed JSONL, manifests, evidence, and reconstruction code.