|
--- |
|
license: apache-2.0 |
|
language: |
|
- en |
|
datasets: |
|
- Magpie-Align/Magpie-Reasoning-V1-150K-CoT-QwQ |
|
library_name: transformers |
|
tags: |
|
- text-generation-inference |
|
- transformers |
|
- unsloth |
|
- qwen2 |
|
- trl |
|
base_model: |
|
- NovaSky-AI/Sky-T1-32B-Preview |
|
--- |
|
 |
|
|
|
# Sky-T1-32B-Preview Fine-Tuned Model |
|
|
|
## Model Details |
|
|
|
- **Developed by:** Daemontatox |
|
- **Model type:** Text Generation |
|
- **Language(s):** English |
|
- **License:** Apache 2.0 |
|
- **Finetuned from model:** [NovaSky-AI/Sky-T1-32B-Preview](https://huggingface.co/NovaSky-AI/Sky-T1-32B-Preview) |
|
- **Training dataset:** [Magpie-Reasoning-V1-150K-CoT-QwQ](https://huggingface.co/datasets/Magpie-Align/Magpie-Reasoning-V1-150K-CoT-QwQ) |
|
- **Training framework:** [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's [TRL](https://github.com/huggingface/trl) library |
|
|
|
## Model Description |
|
|
|
This model is a fine-tuned version of the **NovaSky-AI/Sky-T1-32B-Preview** model, specifically optimized for text generation tasks. It was trained on the **Magpie-Reasoning-V1-150K-CoT-QwQ** dataset, which focuses on reasoning and chain-of-thought (CoT) tasks. The training process was accelerated using **Unsloth**, achieving a 2x speedup compared to traditional methods. |
|
|
|
## Intended Use |
|
|
|
This model is intended for **text generation** tasks, particularly those requiring reasoning and logical coherence. It can be used for: |
|
|
|
- Chain-of-thought reasoning |
|
- Question answering |
|
- Content generation |
|
- Educational tools |
|
|
|
## Training Details |
|
|
|
- **Training framework:** Unsloth + Huggingface TRL |
|
- **Training speed:** 2x faster than traditional methods |
|
- **Dataset:** Magpie-Reasoning-V1-150K-CoT-QwQ |
|
- **Base model:** NovaSky-AI/Sky-T1-32B-Preview |
|
|
|
## How to Use |
|
|
|
You can use this model with the Huggingface `transformers` library: |
|
|
|
```python |
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
# Load the model and tokenizer |
|
model = AutoModelForCausalLM.from_pretrained("Daemontatox/Sky-T1-32B-Preview-Finetuned") |
|
tokenizer = AutoTokenizer.from_pretrained("Daemontatox/Sky-T1-32B-Preview-Finetuned") |
|
|
|
# Generate text |
|
input_text = "Explain the concept of chain-of-thought reasoning." |
|
inputs = tokenizer(input_text, return_tensors="pt") |
|
outputs = model.generate(**inputs, max_length=200) |
|
|
|
# Decode and print the output |
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True)) |
|
``` |
|
|
|
## Limitations |
|
-**The model may generate incorrect or nonsensical responses if the input is ambiguous or outside its training domain.** |
|
|
|
-**It is primarily trained on English data, so performance may degrade for other languages.** |
|
|
|
## Ethical Considerations |
|
-**Bias:** The model may inherit biases present in the training data. Users should be cautious when deploying it in sensitive applications. |
|
|
|
-**Misuse:** The model should not be used for generating harmful, misleading, or unethical content. |
|
|
|
|
|
|
|
|
|
``` |
|
@misc{novasky-sky-t1-32b-preview, |
|
author = {NovaSky-AI}, |
|
title = {Sky-T1-32B-Preview}, |
|
year = {2023}, |
|
publisher = {Hugging Face}, |
|
howpublished = {\url{https://huggingface.co/NovaSky-AI/Sky-T1-32B-Preview}}, |
|
} |
|
|
|
@misc{unsloth, |
|
author = {Unsloth Team}, |
|
title = {Unsloth: Faster Training for Transformers}, |
|
year = {2023}, |
|
publisher = {GitHub}, |
|
howpublished = {\url{https://github.com/unslothai/unsloth}}, |
|
} |
|
|
|
``` |
|
|
|
|
|
## Acknowledgements |
|
Thanks to **NovaSky-AI** for the base model. |
|
|
|
Thanks to **Unsloth** for the faster training framework. |
|
|
|
Thanks to **Huggingface** for the TRL library and tools. |
