LoRA adapter for unsloth/Qwen3.5-2B-Base that detects gender-biased spans in Thai social-media text. Given Thai input, the model returns the same text with bias spans wrapped in inline tags:

<GB-ATTACK>...</GB-ATTACK> — gender-based attacks, insults, dehumanization
<GB-NORMATIVE>...</GB-NORMATIVE> — stereotypes, gender-role policing
<GB-SEX>...</GB-SEX> — sexualized degradation tied to gender

Text without gender bias is returned unchanged.

Training

Fine-tuned with Unsloth on annotated + synthetic Thai examples: thunder121x/thai-gender-bias-training-data

Source code and pipeline: github.com/thunder121x/gender-bias-detection (see services/lora_finetuning/ for the training and inference scripts, including the exact system prompt required at inference time).

Sequence classifier: thunder121x/thai-gender-bias-classifier-minilm

Authors

Nanphat Tongsirisukool — nanphatx@hotmail.com
Natcha Trairattanasak — pnbookclub@gmail.com

thunder121x/thai-gender-bias-span-extraction-qwen3.5-2b

Thai Gender Bias Span Extraction — Qwen 3.5 2B (LoRA)

Training

Related

Authors

thunder121x/
thai-gender-bias-span-extraction-qwen3.5-2b