---
library_name: transformers
tags:
- text-generation
- casual-lm
- sft
- trl
base_model: Qwen/Qwen2.5-3B
datasets:
- HuggingFaceTB/smoltalk
language:
- en
pipeline_tag: text-generation
---

# Qwen-2.5-3B Smoltalk SFT

This is a fine-tuned version of the 3-billion parameter **Qwen/Qwen2.5-3B** base model. It has been instruction fine-tuned via Low-Rank Adaptation (LoRA) and fully merged.

## Model Details

- **Base Model:** Qwen/Qwen2.5-3B
- **Fine-tuning Dataset:** HuggingFaceTB/smoltalk (everyday-conversations subset)
- **Methodology:** Supervised Fine-Tuning (SFT) using TRL
- **Hardware Used:** 1 x NVIDIA L4 GPU (24GB VRAM)

## How to Get Started

You can load and use this model directly with the Hugging Face `pipeline` API.

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

MODEL_ID = "Kerassy/qwen-2.5-3b-smoltalk-sft"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    torch_dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16,
    device_map="auto"
)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

messages = [
    {"role": "user", "content": "Why is the sky blue?"}
]

formatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

outputs = pipe(
    formatted_prompt, 
    max_new_tokens=128, 
    do_sample=True, 
    temperature=0.7,
    top_k=40,
    clean_up_tokenization_spaces=False,
    pad_token_id=tokenizer.pad_token_id,
    eos_token_id=tokenizer.encode("<|end|>")[0] if "<|end|>" in tokenizer.get_vocab() else tokenizer.eos_token_id
)

print(outputs[0]['generated_text'])