snake4u1/strisakhi-gemma4-lora

StriSakhi-Gemma4-2B-LoRA — Legal AI Advocate for Indian Women

Developed by: Shubendu Biswas Competition: Kaggle Gemma 4 Hackathon / Women in AI Challenge
Base Model: unsloth/gemma-4-E2B-it-unsloth-bnb-4bit
Model Type: Causal Language Model (LoRA Adapter)
Languages: Hindi (Devanagari), English, Hinglish input → Hindi/English output
License: Apache-2.0 (same as base)
Finetuned by: Unsloth + PEFT (Hugging Face ecosystem)

---

Model Summary

StriSakhi ("Legal Companion") is a fine-tuned Gemma-4 2B Instruct model specialized as a warm, authoritative legal guide for Indian women seeking rights-based information. Unlike general-purpose LLMs, it is explicitly trained to:

• Respond in simple, sister-like Hindi (or English when requested)
• Cite actual Indian laws with correct section numbers (DV Act 2005, POSH Act 2013, CrPC 125, etc.)
• Structure every response into 5 mandatory blocks: Empathy → Rights → Action Timeline → Helpline → Follow-up Question
• Maintain ≥85% Devanagari purity for Hindi sessions (no Roman script leakage)
• Refuse to generate harmful advice (e.g., never suggests "compromise" in domestic violence cases)

Key Differentiator: This is a safety-first, rights-first legal domain model with structured output conditioning baked into the LoRA weights via 549 curated conversational examples.

---

Competition Results

Benchmark	Score	Pass Rate
Overall (50 cases)	86.4%	43/50 (86%)
Domestic Violence (10)	91.2%	9/10
Property Rights (8)	84.5%	7/8
Maintenance/Divorce (8)	82.1%	6/8
Dowry Harassment (5)	88.0%	4/5
Workplace/POSH (5)	90.0%	5/5
Hinglish → Hindi (8)	85.4%	7/8
Follow-up Short (6)	79.2%	5/6

Benchmark: Custom 50-case legal evaluation suite covering 7 crime categories with automated checks for citation accuracy, Hindi purity, timeline structure, and hallucination resistance.

---

Intended Use

Primary Use Cases

• Legal intake chatbot for NGOs and legal aid clinics serving women in India
• First-response information for domestic violence, property rights, maintenance, dowry, and workplace harassment queries
• Hinglish-to-Hindi translation with legal domain expertise (critical for Tier-2/3 India users)
• Follow-up Q&A after initial legal guidance (short-form answers)

Out-of-Scope Use

• Not a substitute for a licensed advocate. Always directs users to NALSA (15100) and DLSA for actual representation.
• Not for emergency response. Critical emergencies ("happening right now") are handled by a separate hardcoded detector upstream.
• Not for non-Indian jurisdictions. Law citations are India-specific.
• Not for document drafting. Provides guidance, not executable legal documents.

---

Training Details

Hardware

Spec	Value
GPU	NVIDIA Tesla T4 (Kaggle)
VRAM	14.5 GB
Training Time	~35 minutes
Framework	Unsloth 2026.5.2 + Transformers 5.5.0

Hyperparameters

Parameter	Value
Base Model	unsloth/gemma-4-E2B-it-unsloth-bnb-4bit
Method	LoRA (PEFT)
Rank (r)	8
Alpha (lora_alpha)	8
Dropout	0.0
Target Modules	Attention + MLP (vision frozen)
Sequence Length	4096
Quantization	4-bit BnB (NF4)
Batch Size	2
Gradient Accumulation	4
Effective Batch Size	8
Learning Rate	2e-4
LR Scheduler	Linear
Warmup Steps	5
Epochs	3
Optimizer	AdamW 8-bit
Weight Decay	0.001
Seed	42

Dataset

• Size: 549 conversational examples
• Format: ShareGPT-style JSONL with conversations array (system/user/assistant turns)
• Coverage:
  • Domestic Violence (DV Act 2005) — 35%
  • Property / Inheritance — 20%
  • Maintenance / Divorce — 20%
  • Dowry / 498A — 10%
  • Workplace / POSH Act — 10%
  • Follow-up short answers — 5%
• Language Distribution: 70% Hindi output, 20% English output, 10% Hinglish input → Hindi output
• Data Source: Synthetic + manually curated legal scenarios based on actual case patterns from Indian district courts. No private user data.

Training Procedure

1. Template Alignment: Applied Gemma-4 non-thinking chat template to match production llama-server deployment
2. Label Masking: System + user tokens masked as -100 (ignored in loss); only assistant responses trained
3. BOS Deduplication: Removed duplicate <bos> tokens introduced by processor
4. Marker-Based Splitting: Used <|turn>model\n boundary to precisely mask prefix vs. suffix
5. Checkpointing: Saved every 50 steps; best checkpoint at step 207 (epoch 3, final loss: 0.3487)

---

Ethical Statement & Safety

Bias Mitigation

• Gender-specific by design: Model is explicitly conditioned to advocate for women's legal rights; it does not attempt "neutral" framing that could minimize violence (e.g., refuses to call DV a "family matter").
• Language equity: Trained to serve Hinglish-speaking users (common in rural India) by converting to pure Devanagari, reducing the digital language divide.
• Caste/religion awareness: Examples include Hindu Succession Act, Muslim Women Protection Act, and CrPC (secular), avoiding majority-religion bias.

Safety Evaluations

Risk	Mitigation	Status
Hallucinated section numbers	RAG context injected in system prompt; model trained ONLY on provided legal text	Tested
Victim-blaming	Explicit negative training: never says "talk to husband", "compromise", "family matter"	Tested
Emergency mishandling	Upstream hardcoded detector bypasses LLM for active violence; this model handles post-emergency guidance	Tested
Hindi-English script mixing	Purity checker enforces ≥85% Devanagari; LoRA trained on pure Devanagari targets	Tested
Malevolent use (evasion advice)	Refuses to provide advice on evading law; always directs to legal aid	Monitored

Known Limitations

1. RAG dependency: Citation accuracy depends on the quality of retrieved chunks from ChromaDB. Without RAG, the model may hallucinate sections.
2. Thin coverage: Hindu Succession Act, CrPC 125, and Hindu Marriage Act chunks are smaller than DV Act / POSH Act in the retrieval corpus.
3. Token length: Hindi Devanagari consumes ~1.5× tokens per word vs. English; max 4096 context can truncate long RAG contexts.
4. LoRA capacity: Rank-8 is lightweight; complex multi-act reasoning may require full fine-tune or higher rank.

---

How to Use

Quick Inference (Unsloth — recommended)

from unsloth import FastModel
from unsloth.chat_templates import get_chat_template

# Load base + LoRA adapter
model, tokenizer = FastModel.from_pretrained(
    model_name="unsloth/gemma-4-E2B-it-unsloth-bnb-4bit",
    adapter_name="your-hf-username/stri-sakhi-gemma4-2b-lora",  # this repo
    max_seq_length=4096,
    load_in_4bit=True,
)

tokenizer = get_chat_template(tokenizer, chat_template="gemma-4")

messages = [
    {"role": "system", "content": "Tum Kanoon Sakhi ho. Sirf Devanagari Hindi mein jawab do."},
    {"role": "user", "content": "mere pati ne mujhe ghar se nikaala hai"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to("cuda")

outputs = model.generate(
    input_ids=inputs,
    max_new_tokens=512,
    temperature=0.2,
    top_p=0.9,
)
print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])

Merge & Export for Production (llama.cpp / vLLM)

# Merge LoRA into base for single-file deployment
model.save_pretrained_merged(
    "stri-sakhi-merged",
    tokenizer,
    save_method="merged_16bit",  # or "merged_4bit_for_mlx"
)

# Or export to GGUF for llama.cpp server
model.save_pretrained_gguf(
    "stri-sakhi-q4_k_m",
    tokenizer,
    quantization_method="q4_k_m",
)

---

Repository Structure

.
├── README.md                 # This file
├── adapter_config.json       # LoRA config (PEFT)
├── adapter_model.safetensors # LoRA weights (~16 MB)
├── tokenizer/              # Tokenizer files (if customized)
├── benchmark_results.json  # 50-case evaluation raw results
├── training_logs.txt       # Loss curves per step
└── sample_inference.ipynb  # Reproducible inference demo

---

Training Loss Curve

Step	Loss
10	2.373
50	0.315
100	0.162
150	0.130
200	0.123
207 (final)	0.349*

Final epoch loss is higher than mid-epoch because the last batch contains harder, longer examples (property rights with multiple citations).

---

Acknowledgements

• Google DeepMind for the Gemma-4 model family and open weights
• Unsloth team for 2× faster, 50% memory-reduced fine-tuning
• Hugging Face PEFT & Transformers libraries
• Kaggle for Tesla T4 GPU access
• NALSA & DLSA India for the legal aid framework this model promotes

---

Citation

If you use this model in research or production, please cite:

@misc{stri-sakhi-gemma4-2b-2026,
  title = {StriSakhi: A Safety-First Legal Advocate LLM for Indian Women},
  author = {shubendu biswas},
  year = {2026},
  howpublished = {\url{https://huggingface.co/your-username/stri-sakhi-gemma4-2b-lora}},
  note = {Fine-tuned Gemma-4 2B Instruct with LoRA for structured legal guidance}
}

Base model citation:

@article{gemma4-2026,
  title={Gemma 4: A family of highly capable multimodal models},
  author={Google DeepMind},
  year={2026}
}

---

Disclaimer

This model provides general legal information only and does not constitute legal advice. It is not a substitute for a licensed advocate. Always contact NALSA 15100 or your District Legal Services Authority (DLSA) for case-specific representation. The developers assume no liability for actions taken based on model outputs.

---

Model card generated for Hugging Face Open Source AI Challenge — Women Safety & Empowerment Track.