README.md

---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen2.5-1.5B-Instruct
library_name: transformers
tags:
- reinforcement-learning
- text-generation-inference
- science
- code
- math
- finance
pipeline_tag: text-generation
---

![R1.png](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/BKHWttLe9Z8hJ-azW0b8i.png)

# **GCIRS-Reasoning-1.5B-R1**

> **GCIRS-Reasoning-1.5B-R1** is a **research-grade reasoning model** fine-tuned from **Qwen2.5-1.5B-Instruct**, focused on **non-fictional reasoning**, **factual consistency**, and **scientific depth**. Trained with reinforcement learning using the **Big Reasoning Traces** dataset from DeepSeek, this model is tailored for complex analytical tasks and scientific rigor in high-stakes or research environments.

> \[!note]
> GGUF: [https://huggingface.co/prithivMLmods/GCIRS-Reasoning-1.5B-R1-GGUF](https://huggingface.co/prithivMLmods/GCIRS-Reasoning-1.5B-R1-GGUF)

---

## **Key Features**

1. **Reinforcement Learning on Big Reasoning Traces**
   Fine-tuned using **DeepSeek’s Big Reasoning Traces**, ensuring clarity in multi-step reasoning, factual deduction, and long-form scientific argumentation.

2. **Research-Ready Scientific Fidelity**
   Designed for researchers, educators, and analysts—offers **reliable factual recall**, **logical structuring**, and precise step-by-step explanation.

3. **Structured Output in LaTeX, Markdown, and JSON**
   Supports technical documentation and publishing with seamless integration of **LaTeX equations**, **Markdown formatting**, and **JSON output**.

4. **Multilingual Technical Reasoning**
   Effective across **20+ languages**, especially in **scientific**, **academic**, and **technical domains**.

5. **Efficient for Inference**
   Despite its **1.5B parameter scale**, it's optimized for **low-latency inference** across **modern GPUs** and **research pipelines**.

---

## **Quickstart with Transformers**

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/GCIRS-Reasoning-1.5B-R1"

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

prompt = "Explain the principle of entropy in thermodynamics with examples."

messages = [
    {"role": "system", "content": "You are a scientific reasoning assistant."},
    {"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**

* Scientific and research-grade question answering
* Conceptual explanations in physics, biology, and chemistry
* Factual, non-fictional structured content generation
* Academic tutoring and reasoning assessment
* High-fidelity inference in low-latency research settings

## **Limitations**

* Not designed for casual chat or storytelling
* Performance may decline outside scientific/technical domains
* Limited creativity and abstract generalization
* Context limitations in extremely long research documents

## **References**

1. [Qwen2.5 Technical Report (2024)](https://arxiv.org/pdf/2412.15115)
2. [Big Reasoning Traces (DeepSeek Research)]()
3. [Reinforcement Learning with Human Feedback (RLHF)](https://arxiv.org/abs/1906.01749)