LIMO-Qwen3-8B-Math-Full-Precision-v2

Full-precision (bfloat16) merged model trained with LIMO methodology for mathematical reasoning.

Model Details

Base Model: Qwen/Qwen3-8B (full precision)
Training Method: LIMO (Less is More for Reasoning)
Dataset: GAIR/LIMO (817 high-quality samples)
Training Approach: LoRA fine-tuning → Full merge
Model Size: ~16GB (full precision)
Precision: bfloat16

Training Configuration

LoRA Rank: 8 (conservative)
Learning Rate: 5e-6 (conservative)
Epochs: 1 (prevents overfitting)
Batch Size: 2 (memory optimized)
Gradient Accumulation: 4

Expected Performance

Elementary Math: +2-4% improvement over base Qwen3-8B
High School Math: +2-4% improvement over base Qwen3-8B
Reasoning Quality: Enhanced step-by-step mathematical reasoning

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "Cbgcbg/limo-qwen3-8b-math-full-precision_v2",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(
    "Cbgcbg/limo-qwen3-8b-math-full-precision_v2",
    trust_remote_code=True
)

# Example usage
messages = [
    {"role": "user", "content": "Solve: 2x + 3 = 11"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        do_sample=True
    )

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

Model Comparison

Model	Size	Precision	Performance
Original Gasing-8B	15.26 GB	Full	✅ Baseline
Previous LIMO (quantized)	5.55 GB	4-bit	❌ Degraded
This LIMO (full precision)	~16 GB	bfloat16	✅ Expected +2-4%

Training Details

This model was trained using the LIMO methodology, which demonstrates that high-quality mathematical reasoning can be achieved with minimal but carefully curated training data (817 samples vs typical 100k+ datasets).

Key improvements:

✅ Perfect format alignment with evaluation benchmarks
✅ Conservative training parameters to prevent overfitting
✅ Full precision training and inference
✅ High-quality reasoning chains from GAIR/LIMO dataset

Citation

If you use this model, please cite the LIMO paper:

@misc{ye2025limo,
    title={LIMO: Less is More for Reasoning},
    author={Yixin Ye and Zhen Huang and Yang Xiao and Ethan Chern and Shijie Xia and Pengfei Liu},
    year={2025},
    eprint={2502.03387},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}