---
base_model: NewstaR/Fizik-0.6B-Preview
tags:
- text-generation-inference
- transformers
- unsloth
- qwen3
- trl
- sft
license: apache-2.0
language:
- en
---
# Model: Fizik-0.6B-Pro
Note:
In rare cases, the model may use a different ```<think>``` tag format. This does not affect performance or output quality. We're aware of the issue and are working on a fix.
---

## Description

`Fizik-0.6B-Pro` is a refined reasoning model trained on the `Fizik-SFT-Reasoning` dataset — 11,000 examples of structured, step-by-step thinking. Every sample is tagged with `<think>...</think>`, and **all non-reasoning content was removed**.

This model was built to fix the core issue in the Fizik Preview version: inconsistent reasoning behavior. Now, reasoning is always active when prompted, with no ambiguity.

---

## Behavior

- **Always reasons when prompted**  
  The model consistently follows the `<think>` structure without skipping steps.

- **No fallback to non-reasoning answers**  
  Reasoning is treated as the default behavior.

- **Performs well on multi-step tasks**  
  Especially in areas like math, logic, and multi-hop QA.

---

## Intended Use

- Tasks that require explicit reasoning  
- Safe deployment where reliable logic is needed  
- Research on controlled thought generation

---

## Limitations

- Will **not respond naturally** to prompts that expect short or intuitive answers.  
- Use `Fizik-0.6B-Full` if you need toggleable reasoning behavior.

---