---
tags:
- text-generation-inference
- transformers
- qwen2
license: apache-2.0
language:
- en
---

# sf-100

Conversational fine-tune of Qwen2.5-7B-Instruct, supervised-fine-tuned Hugging Face TRL.

## Model details
- **Architecture:** Qwen2 (7B)
- **Parameters:** ~7.6B (reported as 8B in repo metadata)
- **Precision:** BF16 merged weights, trained on top of a 4-bit bnb-quantized base
- **License:** Apache-2.0
- **Language:** Multi
- **Developed by:** liamka

## Intended use

General-purpose conversational assistant — single- and multi-turn chat.

**Not suitable for** safety-critical settings (medical, legal, financial advice), non-English input (not evaluated), or
high-stakes factual lookup without verification.

## Usage

### Transformers

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

model_id = "liamka/sf-100"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Hi, who are you?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

out = model.generate(inputs, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.9)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
```

## Training

- **Framework:** Unsloth + TRL (`SFTTrainer`)
- **Method:** Supervised fine-tuning on top of Qwen2.5-7B-Instruct
- **No RLHF / DPO** applied

Dataset, step count and hyperparameters are not published.

## Limitations

- Inherits biases and knowledge cutoff from Qwen2.5-7B-Instruct.
- SFT only — no preference optimisation, so safety and refusal behaviour matches the base or weaker.
- Can hallucinate. Verify factual claims.
- Evaluated only informally; no benchmark numbers reported.

## Acknowledgements

- [Qwen2.5](https://huggingface.co/Qwen) — Alibaba
- [TRL](https://github.com/huggingface/trl) — Hugging Face