Model Card: Parveshiiii/M1-MathX

Model Details

• Model Name: Parveshiiii/M1-MathX
• Base Architecture: Gemma (1B parameters)
• Model Type: Causal Language Model (text-generation)
• Training Framework: Hugging Face Transformers
• Precision: fp16
• Attention Mechanism: Hybrid sliding-window and full attention layers
• Tokenizer: Gemma tokenizer (vocab size 262,144)

Usage

from transformers import pipeline, TextStreamer

pipe = pipeline("text-generation", model="Parveshiiii/M1-MathX")
messages = [
    {"role": "user", "content": "Who are you?"},
]
streamer = TextStreamer(pipe.tokenizer)
pipe(messages, streamer=streamer, max_new_tokens=10000)

Intended Use

• Designed for mathematical reasoning tasks, including problem solving, equation manipulation, and step-by-step derivations.
• Suitable for educational contexts, math tutoring, and research experiments in reasoning alignment.
• Not intended for general-purpose conversation or sensitive domains outside mathematics.

Training Data

• Dataset: MathX (curated mathematical reasoning dataset)
• Samples Used: ~300
• Training Steps: 50
• Method: GRPO (Group Relative Policy Optimization) fine-tuning
• Objective: Reinforcement-style alignment for improved reasoning clarity and correctness.

Performance

• Demonstrated strong performance on small-scale math problems and symbolic reasoning tasks.
• Early benchmarks suggest improved accuracy compared to the base Gemma 1B model on math-specific datasets.
• Requires formal evaluation on GSM8K, MATH, and other benchmarks for quantitative comparison.

Limitations

• Small dataset and limited training steps mean coverage is narrow.
• May overfit to MathX patterns and fail on broader or more complex problems.
• Not guaranteed to generalize outside mathematical reasoning.
• As a 1B model, capacity is limited compared to larger LLMs.

Ethical Considerations

• Intended for safe educational use.
• Should not be deployed in high-stakes environments without further validation.
• Outputs may contain errors; human oversight is required.

Citation

If you use this model, please cite as:

@misc{Parvesh2025M1MathX,
  author = {Parvesh Rawal},
  title = {Parveshiiii/M1-MathX: A Gemma-1B model fine-tuned on MathX with GRPO},
  year = {2025},
  howpublished = {\url{https://huggingface.co/Parveshiiii/M1-MathX}}
}

---