deepseek-math-7b-3epoch-678k-fullft/README.md

---
language: en
license: apache-2.0
library_name: transformers
base_model: deepseek-ai/deepseek-math-7b-base
tags:
  - mathematics
  - math
  - reasoning
  - deepseek
  - fine-tuned
  - llama
  - causal-lm
  - competition-math
  - gsm8k
  - numinamath
  - metamath
  - aime
  - amc
  - olympiad
datasets:
  - AI-MO/NuminaMath-CoT
  - meta-math/MetaMathQA
  - HuggingFaceH4/MATH-Train
metrics:
  - accuracy
  - perplexity
model-index:
  - name: deepseek-math-7b-3epoch-678k-fullft
    results:
      - task:
          type: text-generation
          name: Mathematical Reasoning
        dataset:
          type: gsm8k
          name: GSM8K
        metrics:
          - type: accuracy
            value: 82.0
pipeline_tag: text-generation
inference: true
---

# 🚀 DeepSeek Math 7B - Full Fine-tune

[![GSM8K](https://img.shields.io/badge/GSM8K-82.0%25-brightgreen)](.)
[![Open LLM Leaderboard](https://img.shields.io/badge/Open%20LLM%20Leaderboard-Submitted-orange)](.)

**Model ID:** `sid172002/deepseek-math-7b-3epoch-678k-fullft`

Fully fine-tuned DeepSeek Math 7B on 678K high-quality math problems. **+17.8% improvement** over base model on GSM8K (64.2% → 82.0%).

## 📊 Quick Stats

| Metric | Value |
|--------|-------|
| Parameters | 7 Billion |
| Training Steps | 63,609 |
| Epochs | 3.0 |
| Dataset Size | 678,494 samples |
| GSM8K Score | **82.0%** |
| Training Loss | 0.6394 |
| Eval Loss | 0.6411 |

## 🏆 Benchmarks

| Benchmark | Score | Base | Improvement |
|-----------|-------|------|-------------|
| **GSM8K** | **82.0%** | 64.2% | **+17.8%** |
| **MathBench** | **92.0%** | ~70% | **+22%** |
| **MMLU** | Pending | - | Leaderboard Eval |
| **MATH** | TBD | 33.2% | In Progress |

## 💻 Quick Start

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "sid172002/deepseek-math-7b-3epoch-678k-fullft",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
    "sid172002/deepseek-math-7b-3epoch-678k-fullft",
    trust_remote_code=True
)

# Solve math problem
problem = "What is 2x + 5 = 13?"
prompt = f"Solve step by step:\n\n{problem}\n\nSolution:\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))