Social Risk Qwen Jev

No. Jev is used as an external hosted decision service under a fixed policy; the repository trains only the Qwen3-VL LoRA that produces evidence fields, and Qwen's final label is removed before the Jev call.

No. The low, medium and high bands are operational segments; the README states they are not calibrated true-risk probabilities and that uncertain outputs should go to human review.

It does not win. On the frozen 170-item test set, TF-IDF with Logistic Regression at C=4 tops the leaderboard at 90.36% Macro-F1, ahead of the preselected text-OR-multimodal plus Jev plan at 87.25% and the LoRA evidence plus Jev arm at 79.74%. The README keeps that negative result visible and says the numbers are not a SOTA claim on 170 items.

At the same t=0.40 threshold, moving from pure Jev text to LoRA image-plus-text evidence plus Jev takes Macro-F1 from 63.04% to 79.74%, harmful recall from 25.53% to 68.09%, false negatives from 35 to 15, and false positives from 2 to 12. The README is careful that the gain mixes in Qwen's semantic processing, so it is not attributable to images alone.

LT-EDI Chinese misogyny meme data with images, transcripts and a binary label, taken from a fixed upstream commit and split train 978, validation 172, test 170, with duplicate candidates kept out of overlapping splits. One multi-frame image and 39 deduplication or leakage items were quarantined without touching the originals. The README stresses that the original dev 170 became the test set, so this is not an official hidden test set.

Downstream receives the Chinese transcript and six evidence fields only, never images, ground-truth labels, Qwen's final label or sample paths, after identifier removal and basic redaction that does not guarantee all personal data is gone. Offline, Python 3.12 with pip install -e ".[cpu]" gives you social-risk results --split test and social-risk demo, which read no API key, call no network and load no model; demo uses synthetic content and preset scores, and real re-runs need GPU and cloud access.

Base is Qwen3-VL-32B-Instruct quantized to 4-bit with bitsandbytes, with LoRA on the language attention q, k, v and o projections at rank 8, alpha 16 and dropout 0.05, leaving the vision encoder frozen at roughly 19.92M trainable parameters. One epoch over 245 optimizer steps, micro-batch 1 with gradient accumulation 4, learning rate 1e-5 decaying to 1e-6 after 12 warmup steps and gradient clipping at 1.0, on a single RTX PRO 6000 Blackwell 96GB.