---
base_model: unsloth/qwen3-0.6b
tags:
- text-generation-inference
- transformers
- unsloth
- qwen3
- trl
license: apache-2.0
language:
- en
datasets:
- open-thoughts/OpenThoughts2-1M
library_name: transformers
---

![Grifflet Logo](./image.png)

# Grifflet-0.6B

**Developed by:** Daemontatox
**License:** Apache-2.0
**Base Model:** [Daemontatox/Grifflet-0.6B](https://huggingface.co/Daemontatox/Grifflet-0.6B)

## Model Overview

Grifflet-0.6B is a lightweight, fine-tuned transformer model designed for efficient **reasoning**, **math problem-solving**, and **code generation**. Despite its small size (600 million parameters), it delivers strong performance for structured tasks requiring logical coherence, step-by-step thinking, and multi-turn conversations.

This model is optimized using TRL and LoRA with Unsloth acceleration for improved speed and memory efficiency.

## Training Dataset

* **Dataset:** [OpenThoughts2-1M](https://huggingface.co/datasets/open-thoughts/OpenThoughts2-1M)
* **Size:** \~1.1M high-quality samples
* **Content Focus:** Stepwise reasoning, logic puzzles, math proofs, structured code generation, educational conversations
* **Tools:** [Curator Viewer](https://curator.bespokelabs.ai/)

The dataset builds on OpenThoughts-114k and incorporates samples from OpenR1-Math, KodCode, and other logic-focused corpora.

## Intended Use Cases

* Educational chatbots for math and programming
* AI agents requiring clear step-by-step reasoning
* Code generation tools for simple to intermediate logic
* Lightweight deployments on resource-constrained hardware

## Known Limitations

* Primarily trained on English; limited multilingual support
* May hallucinate or generate incorrect factual content
* Performance may decline on abstract or high-complexity queries due to model size

## Quick Example

```python
from transformers import pipeline

pipe = pipeline("text-generation", model="Daemontatox/Grifflet-0.6B")
response = pipe("What is the derivative of x^2?")
print(response[0]['generated_text'])
```

## Technical Training Details

* **Framework:** TRL + LoRA with Unsloth acceleration
* **Training Volume:** \~1M samples
* **Hardware:** A100 80GB or equivalent GPU
* **Objective:** Enable coherent, structured reasoning under constrained compute budgets

---