Modelpublic

tojpaj/legal-aid-routing-model-multilingual

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes3downloads

Model Card

Legal Aid Intake Routing — Multilingual (EN/HI/PA)

LoRA adapter for google/gemma-4-31B-it, fine-tuned for NGO legal-aid intake routing across English, Hindi, and Punjabi (Gurmukhi). Given a client's plain-language legal problem in any of these three languages, the model outputs the relevant practice area (16 classes) plus a short preliminary analysis.

Built for the Adaption AutoScientist Challenge (legal category) multilingual bonus track.

Training data

2,898 rows: 1,000 English + 1,000 Hindi + 898 Punjabi, all sourced from the AutoScientist-adapted enhanced_prompt (not raw pre-adaptation text) — tojpaj/legal-aid-intake-routing.

Hindi: machine-translated via MarianMT (Helsinki-NLP/opus-mt-en-hi). Punjabi: machine-translated via Gemma 4 (google/gemma-4-31B-it) through Together AI, Gurmukhi script.

Honest caveat: this translates US legal-aid content (LegalBench / LearnedHands) into Hindi and Punjabi — it demonstrates multilingual routing on the same 16-class taxonomy, but is not a native Indian-law corpus. A genuinely India-native version would need separately-sourced Hindi/Punjabi legal-aid data under Indian law.

Training details

—Base: google/gemma-4-31B-it
—Method: LoRA (r=8, alpha=16), SFT, via Together AI
—3 epochs, learning_rate=2e-4
—Fine-tuning job: ft-65a89869-3409

Usage

This is a LoRA adapter, not a merged model — load on top of the base model, or serve via Together's dedicated endpoint using the fine-tuned model name tojpaj_33d2/gemma-4-31B-it-legal-aid-multilingual-2321b6bb.

Framework versions

—PEFT 0.15.1