gemma-2-2b-SFT-Reasoning-full-Model/README.md

---
base_model: google/gemma-2-2b
library_name: transformers
tags:
  - math
  - reasoning
  - sft
  - gemma
datasets:
  - nohurry/Opus-4.6-Reasoning-3000x-filtered
language:
  - en
  - th
license: gemma
---

# Gemma-2-2B SFT Reasoning Model

A supervised fine-tuned version of [`google/gemma-2-2b`](https://huggingface.co/google/gemma-2-2b), trained to produce structured chain-of-thought reasoning on mathematical and logical problems.

> **Model Lineage:** This SFT model serves as the foundation for downstream training:
> [**Phonsiri/gemma-2-2b-GRPO-Reasoning-full**](https://huggingface.co/Phonsiri/gemma-2-2b-GRPO-Reasoning-full)

---

## Model Highlights

This model was trained to explicitly separate its reasoning process from its final answer, using a structured output format. It learns the **syntax and structure** of chain-of-thought reasoning before any reinforcement signal is applied.

**Output Format:**

| Section | Tag | Description |
|---|---|---|
| Chain-of-Thought | `<reasoning> ... </reasoning>` | Step-by-step internal reasoning |
| Final Answer | `<answer> ... </answer>` | Concise final answer |

---

## Training Details

### Base Model
Fine-tuned from [`google/gemma-2-2b-it`](https://huggingface.co/google/gemma-2-2b-it) using full parameter fine-tuning (no LoRA/PEFT).

### Datasets
Training data was combined from the following sources:

| Dataset | Type |
|---|---|
| [`nohurry/Opus-4.6-Reasoning-3000x-filtered`](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered) | HuggingFace — Reasoning |
| `math_combined_2566_2567.json` | Local — Thai math problems |
| `problems_1_5.json` | Local — Math problems |
| `problems_6_10.json` | Local — Math problems |
| `problems_101_125.json` | Local — Math problems |
| `combined.json` | Local — Combined problems |
| `all_solutions.json` | Local — Solutions |

### Hyperparameters

| Parameter | Value |
|---|---|
| Epochs | 3 |
| Learning Rate | 2e-5 |
| LR Scheduler | Cosine |
| Max Seq Length | 8192 |
| Batch Size (per device) | 4 |
| Gradient Accumulation | 4 |
| Effective Batch Size | 16 |
| Warmup Ratio | 0.1 |
| Weight Decay | 0.01 |
| Precision | bfloat16 |
| Gradient Checkpointing | Yes |
| Attention | SDPA |
| Optimizer | AdamW |

---

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
import torch

model_id = "Phonsiri/gemma-2-2b-SFT-Reasoning-full-Model"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16
)

system_prompt = (
    "You are a helpful assistant. Please reason step by step, "
    "and put your thoughts within <reasoning> and </reasoning> tags, "
    "and your final answer within <answer> and </answer> tags."
)
prompt = "Solve for x: 3x + 5 = 20"

messages = [{"role": "user", "content": f"{system_prompt}\n\n{prompt}"}]

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

streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=False)

with torch.no_grad():
    model.generate(
        **inputs,
        streamer=streamer,
        max_new_tokens=4096,
        temperature=0.6,
        top_p=0.9,
        repetition_penalty=1.1,
    )
```

### Example Output

```
<reasoning>
We need to isolate x on one side of the equation.

Step 1: Subtract 5 from both sides.
  3x + 5 - 5 = 20 - 5
  3x = 15

Step 2: Divide both sides by 3.
  x = 15 / 3
  x = 5
</reasoning>

<answer>
x = 5
</answer>
```

---

## Acknowledgements

**Authors:**
- **Phonsiri Thabunsri** — [@Phonsiriwillbejommarn](https://github.com/Phonsiriwillbejommarn)
- **CYP777** — [@CYP777](https://github.com/CYP777)

**Project Advisor:**
- **Supaporn Bunrit, Ph.D.** — Suranaree University of Technology

**Institutions & Credits:**
- **Suranaree University of Technology (SUT)** — Research support and computing resources
- **Google DeepMind** — Open-weights Gemma 2 model