---
license: apache-2.0
language:
- en
library_name: peft
pipeline_tag: text-generation
tags:
- qwen2
- lora
- peft
- sft
- trl
- transformers
- sakthai
- tool-calling
- instruct
- function-calling
- text-generation
datasets:
- Nanthasit/sakthai-combined-v5
base_model: Qwen/Qwen2.5-7B-Instruct
---

# SakThai Context 7B — LoRA Adapter

A LoRA fine-tune of [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) for structured tool-calling and instruction following, trained on the SakThai tool-calling curriculum.

## Model Details

- **Developed by:** Nanthasit
- **Base model:** Qwen/Qwen2.5-7B-Instruct (7B parameters)
- **Architecture:** Qwen2.5 decoder-only transformer + LoRA adapters
- **Fine-tuning method:** LoRA (rank=16, alpha=32) via TRL SFTTrainer
- **Training data:** [Nanthasit/sakthai-combined-v5](https://huggingface.co/datasets/Nanthasit/sakthai-combined-v5)
- **License:** Apache 2.0
- **Inference:** BF16 (use `transformers` with `device_map="auto"`)

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype="bfloat16",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

model = PeftModel.from_pretrained(base_model, "Nanthasit/sakthai-context-7b-tools")
```

### Chat Template

The model uses Qwen2.5's standard chat template with system/user/assistant roles and supports function-calling via the `tools` parameter in the tokenizer.

## Merged Version

For production inference, use the merged model instead:
👉 [Nanthasit/sakthai-context-7b-merged](https://huggingface.co/Nanthasit/sakthai-context-7b-merged)

## Intended Use

- Tool-calling and function-calling agents
- Structured instruction following
- Chat and assistant applications requiring external tool use

## Training Details

- **Framework:** Hugging Face TRL (SFTTrainer)
- **Compute:** HF Jobs (T4 GPU)
- **Quantization:** 4-bit NF4 for training
- **Dataset size:** ~4,000+ tool-calling examples
- **LoRA config:** `r=16, lora_alpha=32, target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]`