---
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/blob/main/LICENSE
language:
- en
pipeline_tag: text-generation
base_model: Qwen/Qwen2.5-1.5B-Instruct
tags:
- chat
- trl
- sft
- math
library_name: transformers
model-index:
- name: Qwen2.5-1.5B-Instruct-QwQ
  results:
  - task:
      type: text-generation
    dataset:
      name: GSM8k
      type: gsm8k
    metrics:
    - name: pass@4
      type: pass@4
      value: 89.6
      verified: false
---
# Qwen2.5-1.5B-Instruct-QwQ

## Introduction

Qwen2.5-1.5B-Instruct-QwQ is a fine-tuned model based on Qwen2.5-1.5B-Instruct. It was fine-tuned on roughly 20k samples from QwQ-32B-Preview. Compared to Qwen2.5-1.5B-Instruct, this fine-tuned model seems more performant in mathematics contexts and general reasoning. Also it shows some capabilities of self-correction, altough it seems a bit limited (bigger models seem to learn self-correction better, e.g. the 3B & 7B version show much better self-correction abilities in my experiments).

For data generation, math problems from the train sets of the GSM8k and MATH datasets were used.


**This repo contains the instruction-tuned 1.5B Qwen2.5 model fine-tuned on QwQ reasoning chains**, which has the following features:
- Type: Causal Language Models
- Training Stage: Pretraining & Post-training
- Architecture: transformers with RoPE, SwiGLU, RMSNorm, Attention QKV bias and tied word embeddings
- Number of Parameters: 1.54B
- Number of Paramaters (Non-Embedding): 1.31B
- Number of Layers: 28
- Number of Attention Heads (GQA): 12 for Q and 2 for KV
- Context Length: Full 32,768 tokens and generation 8192 tokens
  

## Quickstart

Here provides a code snippet with `apply_chat_template` to show you how to load the tokenizer and model and how to generate contents.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "micaebe/Qwen2.5-1.5B-Instruct-QwQ"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "system", "content": "You are a helpful and harmless assistant. You are Qwen developed by Alibaba. You should think step-by-step."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
```

For GSM8k performance comparison: the base-instruct model scores 73.2% on the test set.

Disclaimer: GSM8k scores are currently only from the first 27% of the test set. Scores will be updated in the future, tested on the full dataset.