themis-mistral-7b-lora

What This Is

A LoRA adapter fine-tuned on 1,939 Indian legal Q&A pairs (BNS 2023, BNSS 2023, IPC, Consumer Protection Act, RTI Act). Must be loaded on top of unsloth/mistral-7b-instruct-v0.3-bnb-4bit.

What This Demonstrates

• End-to-end fine-tuning pipeline: data scraping → synthetic generation → LoRA training (Unsloth/Kaggle T4) → HuggingFace deployment
• Instruction-following behavior in legal assistant style
• Correctly trained disclaimer behavior
• Partially learned response structure (citations, recommendations)

What This Does NOT Demonstrate

• Accurate section number citation (~60% hallucination rate on BNS-specific queries)
• BNS abbreviation recognition (model confuses "BNS" with unrelated expansions)
• Deep statutory knowledge (1,939 pairs was insufficient for domain grounding)
• IPC → BNS section mapping

Root cause: Mistral 7B has near-zero BNS 2023 pretraining knowledge — BNS was enacted Dec 2023, at/after Mistral's training cutoff. The fine-tune taught style, not substance.

---

Training Details

---

How to Use

With PEFT + Transformers

python
Copy code
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
)

tokenizer = AutoTokenizer.from_pretrained("unsloth/mistral-7b-instruct-v0.3-bnb-4bit")
base_model = AutoModelForCausalLM.from_pretrained(
    "unsloth/mistral-7b-instruct-v0.3-bnb-4bit",
    quantization_config=bnb_config,
    device_map="auto",
)

model = PeftModel.from_pretrained(base_model, "Daniel2503/themis-mistral-7b-lora")
model.eval()

# Use FULL ACT NAMES for best results (e.g., "Bharatiya Nyaya Sanhita" not "BNS")
prompt = "### Instruction:\nWhat is the punishment for theft under the Bharatiya Nyaya Sanhita?\n\n### Response:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.3)

response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)

---

Known Limitations (v1)

• Section hallucination — ~60% of BNS-specific queries contain fabricated section numbers
• Abbreviation confusion — "BNS" not recognized; use full name "Bharatiya Nyaya Sanhita"
• Insufficient training data — 1,939 pairs teaches style, not statutory content
• No case law — statutes only, no judgments
• English only — Hindi on roadmap
• State-specific laws not covered

For Best Results

• Use full act names: "Bharatiya Nyaya Sanhita" not "BNS"
• Ask general legal questions, not specific section numbers
• Treat output as orientation, never as authoritative legal reference

---

v2 Roadmap

Success criteria: Model correctly identifies BNS as Bharatiya Nyaya Sanhita and cites accurate section numbers on 70%+ of criminal law queries.

---

Citation

bibtex
Copy code
@misc{themis2026,
  title={THEMIS: Parametric Legal Intelligence Engine for Indian Law},
  author={Daniel Deshmukh},
  year={2026},
  howpublished={\url{https://huggingface.co/Daniel2503/themis-mistral-7b-lora}},
}

---

License

MIT License

---

THEMIS v1 proves the pipeline works. v2 will make the model actually know Indian law.