metadata
datasets:
- XenArcAI/MathX-5M
base_model:
- google/gemma-3-1b-it
pipeline_tag: text-generation
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}}
}