Downloads · 30 days
15
38% of all-time downloads
mr-mc/flowjudge-dialam-qwen3-0.6b-v3-n4096
flowjudge-dialam-qwen3-0.6b-v3-n4096 is a text generation model from mr-mc. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as other.
This is the selected FlowJudge QLoRA adapter for one narrow behavior: given one new proposition and a complete fixed-size block of earlier propositions, emit a bare JSON object containing every and only direct SUPPORT…
Downloads · 30 days
15
38% of all-time downloads
All-time downloads
39
Public
Repo size
51.8 MB
Likes
0
Public
Click a slice to open those files.
.safetensors40.4 MB · 72%
From the Hugging Face model README
This is the selected FlowJudge QLoRA adapter for one narrow behavior: given one new proposition and a complete fixed-size block of earlier propositions, emit a bare JSON object containing every and only direct SUPPORT, ATTACK, or REPHRASE edge to supplied IDs, or an empty relation list.
V3 uses 4,096 private transformed QT30 examples arranged as 2,048 exact same-update positive/NONE pairs. It keeps the v1 Qwen3-0.6B QLoRA optimization configuration and changes the data plus loss reduction: assistant-token loss is averaged within each example before examples are averaged. This corrects the 4.50x positive-to-NONE output-length weighting imbalance diagnosed after v2.
| Metric | Base | v1 n=2048 | v2 n=2048 | Selected v3 n=4096 |
|---|---|---|---|---|
| Exact patch accuracy | 0.0% | 26.7% | 23.3% | 43.3% |
| Edge precision | 0.0% | 16.7% | 18.8% | 43.8% |
| Edge recall | 0.0% | 16.7% | 25.0% | 29.2% |
| Edge F1 | 0.0% | 16.7% | 21.4% | 35.0% |
| Relation macro-F1 | 0.0% | 16.7% | 18.8% | 33.9% |
| False edges/update | 0.000 | 0.667 | 0.867 | 0.300 |
| NONE cases with false edges | n/a | 2/6 | 5/6 | 0/6 |
| Judge Robustness /4 | 1.40 | 2.10 | 1.43 | 2.97 |
V3 cleared its preregistered material-improvement rule and is the selected model direction. It still failed the original high reliability bar: edge F1 is 35.0% versus the required 85%, exact patch accuracy is 43.3% versus 80%, and ATTACK is the weakest relation at 18.2% F1. This is an educational research artifact, not a claim of production-ready argument mining.
The training text is not redistributed. Consumers must obtain the official QT30 corpus and appropriate permission separately. Included evidence contains only aggregate metrics, configuration, hashes, and text-free reconstruction metadata.
Load this PEFT adapter over Qwen/Qwen3-0.6B with a compatible
Transformers/PEFT or Unsloth runtime. Use greedy decoding with Qwen thinking
disabled and the project behavior prompt.