107 lines
3.6 KiB
Markdown
107 lines
3.6 KiB
Markdown
---
|
||
datasets:
|
||
- Magpie-Align/Magpie-Reasoning-V2-250K-CoT-Deepseek-R1-Llama-70B
|
||
license: apache-2.0
|
||
language:
|
||
- en
|
||
base_model:
|
||
- Qwen/Qwen3-1.7B
|
||
pipeline_tag: text-generation
|
||
library_name: transformers
|
||
tags:
|
||
- trl
|
||
- text-generation-inference
|
||
- llama
|
||
- distill
|
||
- experimental
|
||
---
|
||
|
||
|
||
|
||
# **Regulus-Qwen3-R1-Llama-Distill-1.7B**
|
||
|
||
> **Regulus-Qwen3-R1-Llama-Distill-1.7B** is a **distilled reasoning model** fine-tuned on **Qwen/Qwen3-1.7B** using **Magpie-Align/Magpie-Reasoning-V2-250K-CoT-DeepSeek-R1-Llama-70B**.
|
||
> The training leverages **distilled traces from DeepSeek-R1-Llama-70B**, transferring advanced reasoning patterns into a lightweight 1.7B parameter model.
|
||
> It is specialized for **chain-of-thought reasoning across code, math, and science**, optimized for efficiency and mid-resource deployment.
|
||
|
||
> \[!note]
|
||
> GGUF: [https://huggingface.co/prithivMLmods/Regulus-Qwen3-R1-Llama-Distill-1.7B-GGUF](https://huggingface.co/prithivMLmods/Regulus-Qwen3-R1-Llama-Distill-1.7B-GGUF)
|
||
|
||
---
|
||
|
||
## **Key Features**
|
||
|
||
1. **Distilled Reasoning from Large-Scale Models**
|
||
Trained with **distilled traces from DeepSeek-R1-Llama-70B**, preserving structured **chain-of-thought reasoning** in a smaller, faster model.
|
||
|
||
2. **Unified Code + Math + Science Reasoning**
|
||
Strong performance across computational logic, programming tasks, and scientific problem solving.
|
||
|
||
3. **Structured Chain-of-Thought Generation**
|
||
Produces clear, step-by-step explanations for algorithms, equations, and symbolic tasks.
|
||
|
||
4. **Optimized Lightweight Footprint**
|
||
Maintains reasoning depth while being deployable on **mid-range GPUs**, **offline clusters**, and **edge AI systems**.
|
||
|
||
5. **Multi-Format Output Support**
|
||
Generates responses in **LaTeX**, **Markdown**, **JSON**, and **tabular formats** for technical and research workflows.
|
||
|
||
---
|
||
|
||
## **Quickstart with Transformers**
|
||
|
||
```python
|
||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
||
model_name = "prithivMLmods/Regulus-Qwen3-R1-Llama-Distill-1.7B"
|
||
|
||
model = AutoModelForCausalLM.from_pretrained(
|
||
model_name,
|
||
torch_dtype="auto",
|
||
device_map="auto"
|
||
)
|
||
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||
|
||
prompt = "Explain step by step how to solve a system of linear equations using Gaussian elimination."
|
||
|
||
messages = [
|
||
{"role": "system", "content": "You are a reasoning assistant skilled in math, code, and scientific logic."},
|
||
{"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**
|
||
|
||
* **Math and algorithm tutoring** with clear reasoning steps
|
||
* **Code reasoning and synthesis** for debugging and algorithm design
|
||
* **Scientific problem solving** in physics, chemistry, and biology
|
||
* **Structured educational assistant** for step-by-step learning
|
||
* **Efficient deployment** where distilled reasoning fidelity is required
|
||
|
||
## **Limitations**
|
||
|
||
* Derived from **distilled traces** – reasoning may simplify compared to full-scale teacher models
|
||
* Not tuned for general-purpose conversation or creative writing
|
||
* Context length limits multi-document or long-codebase reasoning
|
||
* Optimized for structured reasoning, not emotional or casual dialogue |