Josephgflowers/Tinyllama-r1

This model sponsored by the generous support of Cherry Republic.

https://www.cherryrepublic.com/

Model Overview

TinyLlama-R1 is a lightweight transformer model designed to handle instruction-following and reasoning tasks, particularly in STEM domains. This model was trained using the Magpie Reasoning V2 250K-CoT dataset, with a goal to improve reasoning through high-quality instruction-response pairs. However, based on early tests, TinyLlama-R1 shows reduced responsiveness to system-level instructions, likely due to the absence of system messages in the dataset.

Model Name: Josephgflowers/Tinyllama-STEM-Cinder-Agent-v1

---

Key Features

• Dataset Focus: Built on the Magpie Reasoning V2 250K-CoT dataset, enhancing problem-solving in reasoning-heavy tasks.
• STEM Application: Tailored for tasks involving scientific, mathematical, and logical reasoning.
• Instruction Handling: Initial observations indicate reduced adherence to system instructions, a change from previous versions.

---

Model Details

• Model Type: Transformer-based (TinyLlama architecture)
• Parameter Count: 1.1B
• Context Length: Updated to 8k
• Training Framework: Unsloth
• Primary Use Cases:
  • Inteded for research into COT in small language models
  • Technical problem-solving
  • Instruction-following conversations

---

Training Data

The model was fine-tuned on the Magpie Reasoning V2 250K-CoT dataset. The dataset includes diverse instruction-response pairs, but notably lacks system-level messages, which has impacted the model's ability to consistently follow system directives.

Dataset Characteristics

• Sources:
  Instructions were generated using models like Meta's Llama 3.1 and 3.3.
  Responses were provided by DeepSeek-R1-Distill-Llama-70B.
• Structure: Instruction-response pairs with an emphasis on chain-of-thought (CoT) reasoning styles.
• Limitations: No system-level instructions were included, affecting instruction prioritization and response formatting in some contexts.

---

Known Issues & Limitations

• System Instructions: The model currently does not respond well to system messages, in contrast to previous versions.
• Performance Unverified: This version has not yet been formally tested on benchmarks like GSM-8K.

---

The model can be accessed and fine-tuned via Josephgflowers on Hugging Face.

Training & License Information

License: CC BY-NC 4.0 (Non-commercial use only) This model was trained using datasets under:

Meta Llama 3.1 and 3.3 Community License
CC BY-NC 4.0 (Creative Commons Non-Commercial License)

Acknowledgments

Thanks to the Magpie Reasoning V2 dataset creators and the researchers behind models like Deepseek-R1 and Meta Llama.

@article{xu2024magpie, title={Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing}, author={Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin}, year={2024}, eprint={2406.08464}, archivePrefix={arXiv}, primaryClass={cs.CL} }