Pruning Details

• Pruning method: SparseGPT-style one-shot pruning
• Target sparsity: 30%
• Calibration dataset: mohitskaushal/InLegalLaySum-Phi4-Train-16K
• Calibration samples: 32
• Sequence length used for calibration: 2048
• Pruned modules: decoder Linear layers
• Excluded from pruning: lm_head

Important Note

This model contains zero-valued weights after pruning. The .safetensors file size may not be smaller because dense tensor storage still stores zero values.

Intended Use

This model is intended for research experiments on sparse legal summarization models, especially for comparing dense vs SparseGPT-pruned variants.

Example Use

python
Copy code
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "mohitskaushal/gemma3-legal-sparsegpt-30"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)

prompt = \"\"\"You are a legal summarization assistant. Summarize the following legal text in simple layman language.

Legal text:
The appellant challenged the judgment of the High Court on the ground that the conviction was based on insufficient evidence and that material witnesses were not examined.

Layman summary:\"\"\"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=200,
        do_sample=False,
        repetition_penalty=1.15,
        no_repeat_ngram_size=4,
        pad_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))