LFM2.5-1.2B-JP-math-safe

MATH-optimized | Safe pruning | 30% weights pruned

This model is a conservatively pruned version of LiquidAI/LFM2.5-1.2B-JP.

Note: Minimal quality drop detected. The Wanda pruning algorithm effectively identifies and removes less important weights while preserving model capability.

Performance Comparison

Category	Original	Pruned	Change
Python	0.0%	0.0%	→
Html	15.0%	15.0%	→
Trivia	85.0%	85.0%	→
Math	55.0%	50.0% ⭐	↓ 5.0%
Reasoning	50.0%	45.0%	↓ 5.0%
Medical	70.0%	70.0%	→
Linux	50.0%	50.0%	→
Writing	15.0%	15.0%	→

Average: 42.5% -> 41.2% (-1.2%)

Math Retention: 90.9%

$Comparison Graph$

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("CompactAI/LFM2.5-1.2B-JP-math-safe")
tokenizer = AutoTokenizer.from_pretrained("CompactAI/LFM2.5-1.2B-JP-math-safe")

inputs = tokenizer("Your prompt here", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))