vivekvar/qwen3-14b-pocso-legal-assistant

Qwen3-14B Fine-tuned for Indian Legal Domain (POCSO Investigation Assistant)

Model Description

This is a LoRA fine-tuned version of Qwen/Qwen3-14B specifically trained for Indian legal domain tasks, with a focus on POCSO (Protection of Children from Sexual Offences) Act cases.

Model Details

Property	Value
Base Model	Qwen/Qwen3-14B
Fine-tuning Method	LoRA (Low-Rank Adaptation)
LoRA Rank (r)	64
LoRA Alpha	128
Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training Examples	8,954
Validation Examples	1,054
Epochs	3
Learning Rate	2e-5

Training Data

The model was fine-tuned on Indian legal case examples including:

• POCSO Act cases and sections
• IPC (Indian Penal Code) provisions
• BNS (Bharatiya Nyaya Sanhita) 2023 sections
• CrPC/BNSS procedural guidelines
• Real anonymized complaint documents in English and Telugu

Capabilities

1. SUGGEST_SECTIONS

Suggests applicable legal sections based on complaint text.

2. GENERATE_SUMMARY

Extracts key details (victim, accused, incident) from complaints.

3. IDENTIFY_EVIDENCE

Recommends evidence collection priorities.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-14B",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "vivekvar/qwen3-14b-pocso-legal-assistant")

# Generate
prompt = """Task: Suggest legal sections for this complaint.

Complaint: A 15-year-old girl was harassed by her neighbor.

Output JSON:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation Results

Metric	Score
Semantic Similarity	77.8%
JSON Output Validity	100%
Reliability	100%
Model Grade	EXCELLENT

Comparison with GPT + RAG

Task	Qwen3-14B (this)	GPT + RAG
SUGGEST_SECTIONS	70%	0% (refuses)
GENERATE_SUMMARY	85%	75%
Reliability	100%	66.7%

Limitations

1. May apply wrong law (BNS vs IPC) for cases before July 2024
2. May miss mandatory POCSO sections in edge cases
3. Evidence recommendations less detailed than RAG systems
4. Optimized for English and Telugu

Disclaimer

⚠️ This model is an AI assistant and should NOT replace professional legal advice.

All suggestions must be verified by qualified legal professionals.

License

Apache 2.0

Citation

@misc{qwen3-14b-pocso-legal,
  title={Qwen3-14B Fine-tuned for Indian Legal Domain},
  author={AI4AP Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/vivekvar/qwen3-14b-pocso-legal-assistant}
}