Upload README.md with huggingface_hub

README.md

@@ -1,3 +1,110 @@
 ---
-license: mit
+language:
+- en
+library_name: llama-cpp-python
+tags:
+- legal
+- india
+- gguf
+- quantized
+- rag
+- llama-3.1
+pipeline_tag: text-generation
+base_model: unsloth/Meta-Llama-3.1-8B-Instruct
+license: other
 ---
+
+# LegalSahyak (Q4_K_M GGUF)
+
+## Model Description
+`LegalSahyak_q4_k_m.gguf` is a quantized GGUF model intended for local legal question answering workflows, especially when paired with retrieval over contracts and Indian statutes.
+
+- Base model: `unsloth/Meta-Llama-3.1-8B-Instruct`
+- Adaptation: LoRA fine-tuning (rank `r=128`) and merge
+- Quantization: `q4_k_m` GGUF
+- Primary runtime target: `llama.cpp` / `llama-cpp-python`
+
+## Intended Use
+- Contract clause explanation and extraction
+- Statute-grounded legal QA in a retrieval-augmented (RAG) pipeline
+- Local/offline inference where low memory usage is needed
+
+This model should be used with retrieval and human review for any high-stakes legal scenario.
+
+## Out-of-Scope Use
+- Autonomous legal advice without human oversight
+- Any use requiring guaranteed legal correctness or jurisdictional completeness
+- Sensitive decisions where model hallucinations can cause harm
+
+## Training Data
+The training pipeline in `models/train.py` uses two public datasets:
+
+1. `Prarabdha/indian-legal-supervised-fine-tuning-data`
+2. `opennyaiorg/aalap_instruction_dataset`
+
+Training was performed in two stages:
+
+1. Knowledge injection on legal supervised examples
+2. Behavioral alignment on instruction-following data
+
+## Training Procedure (Summary)
+- Context length: up to `8192`
+- Precision during training: `bfloat16`
+- LoRA target modules: `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
+- Optimizer: `adamw_8bit`
+- Scheduler: cosine
+- Export: merged weights -> GGUF quantized as `q4_k_m`
+
+## Inference
+Example with `llama-cpp-python`:
+
+```python
+from llama_cpp import Llama
+
+llm = Llama(
+    model_path="LegalSahyak_q4_k_m.gguf",
+    n_ctx=4096,
+    n_gpu_layers=20,
+    verbose=False,
+)
+
+resp = llm.create_chat_completion(
+    messages=[
+        {"role": "system", "content": "You are a legal assistant. Use provided context only."},
+        {"role": "user", "content": "Explain the notice period clause in simple words."},
+    ],
+    max_tokens=512,
+    temperature=0.0,
+)
+
+print(resp["choices"][0]["message"]["content"])
+```
+
+## Model File Details
+- Filename: `LegalSahyak_q4_k_m.gguf`
+- Size (bytes): `4920738464`
+- Approx size: `4.58 GiB`
+- SHA256: `F32460DD8E7DC927B3CF33065D1E753FC1F85ED102A678512C8A5F520F544405`
+
+## Limitations
+- Can produce plausible but incorrect legal text
+- Performance depends heavily on retrieval quality and prompt constraints
+- May not reflect the latest statutory amendments
+- Not a substitute for licensed legal counsel
+
+## Bias, Risk, and Safety
+- Dataset and model biases may propagate into outputs
+- Should not be used as the sole basis for legal, compliance, or policy decisions
+- Recommended controls:
+  - Ground responses in retrieved sources
+  - Log model outputs and review manually
+  - Add refusal/uncertainty handling when context is missing
+
+
+
+## Citation
+If you use this model in research or products, cite:
+
+- The base model (`Meta-Llama-3.1`)
+- The datasets listed above
+- This repository (`Legalsahyak`)