110 lines
3.9 KiB
Markdown
110 lines
3.9 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
datasets:
|
||
|
|
- Gen-Verse/ReasonFlux-V2-Reasoner-DPO
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
- zh
|
||
|
|
base_model:
|
||
|
|
- Qwen/Qwen3-1.7B
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
library_name: transformers
|
||
|
|
tags:
|
||
|
|
- trl
|
||
|
|
- text-generation-inference
|
||
|
|
- code
|
||
|
|
- DPO
|
||
|
|
---
|
||
|
|
|
||
|
|
|
||
|
|
|
||
|
|
# **ReasonFlux-Qwen3-dpo**
|
||
|
|
|
||
|
|
> **ReasonFlux-Qwen3-dpo** is a fine-tuned version of **Qwen3-1.7B**, trained on the [**Gen-Verse/ReasonFlux-V2-Reasoner-DPO**](https://huggingface.co/datasets/Gen-Verse/ReasonFlux-V2-Reasoner-DPO) dataset.
|
||
|
|
> It adopts a **template-augmented reasoning paradigm**, internalizing structured **thought templates** through **iterative hierarchical reinforcement learning** and **direct preference optimization (DPO)**.
|
||
|
|
> This design enables the model to reason more transparently, consistently, and adaptively across multi-domain scientific and mathematical tasks.
|
||
|
|
|
||
|
|
> \[!note]
|
||
|
|
> GGUF: [https://huggingface.co/prithivMLmods/ReasonFlux-Qwen3-dpo-GGUF](https://huggingface.co/prithivMLmods/ReasonFlux-Qwen3-dpo-GGUF)
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## **Key Features**
|
||
|
|
|
||
|
|
1. **Template-Augmented Reasoning**
|
||
|
|
Incorporates structured **reasoning templates** that guide step-by-step thinking, improving coherence and reducing hallucinations.
|
||
|
|
|
||
|
|
2. **DPO Fine-Tuning with Hierarchical Reinforcement**
|
||
|
|
Leverages **direct preference optimization** along with **iterative reinforcement learning**, internalizing high-quality reasoning behaviors.
|
||
|
|
|
||
|
|
3. **Scientific & Mathematical Expertise**
|
||
|
|
Excels at symbolic derivations, step-by-step proofs, and multi-domain STEM reasoning (physics, chemistry, biology, mathematics).
|
||
|
|
|
||
|
|
4. **Code Understanding & Generation**
|
||
|
|
Provides detailed coding explanations, debugging support, and optimization hints across multiple programming languages.
|
||
|
|
|
||
|
|
5. **Structured Output Mastery**
|
||
|
|
Fluent in producing outputs across **LaTeX**, **Markdown**, **JSON**, **CSV**, and **YAML** for seamless integration in research and technical workflows.
|
||
|
|
|
||
|
|
6. **Efficient Deployment**
|
||
|
|
Lightweight yet powerful, designed for **mid-range GPUs**, **research clusters**, and **edge AI environments**.
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## **Quickstart with Transformers**
|
||
|
|
|
||
|
|
```python
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
|
||
|
|
model_name = "prithivMLmods/ReasonFlux-Qwen3-dpo"
|
||
|
|
|
||
|
|
model = AutoModelForCausalLM.from_pretrained(
|
||
|
|
model_name,
|
||
|
|
torch_dtype="auto",
|
||
|
|
device_map="auto"
|
||
|
|
)
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
|
|
|
||
|
|
prompt = "Explain how reinforcement learning differs from supervised learning with real-world examples."
|
||
|
|
|
||
|
|
messages = [
|
||
|
|
{"role": "system", "content": "You are a reasoning tutor skilled in science, math, and coding."},
|
||
|
|
{"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]
|
||
|
|
print(response)
|
||
|
|
```
|
||
|
|
|
||
|
|
---
|
||
|
|
|
||
|
|
## **Intended Use**
|
||
|
|
|
||
|
|
* Advanced reasoning tutor for mathematics, coding, and scientific research
|
||
|
|
* Research assistant capable of structured problem-solving with template-guided reasoning
|
||
|
|
* Technical documentation and structured data generation
|
||
|
|
* STEM-focused chatbot or API for research and education workflows
|
||
|
|
* Deployment in environments requiring transparent reasoning with efficient compute use
|
||
|
|
|
||
|
|
## **Limitations**
|
||
|
|
|
||
|
|
* Not optimized for casual or creative writing
|
||
|
|
* Context limitations may restrict multi-document or full codebase comprehension
|
||
|
|
* Specializes in structured reasoning—general chit-chat may underperform
|
||
|
|
* Optimized for **clarity of reasoning** rather than **natural conversational tone**
|