|
|
--- |
|
|
license: mit |
|
|
language: |
|
|
- en |
|
|
base_model: |
|
|
- Qwen/Qwen3-30B-A3B |
|
|
--- |
|
|
|
|
|
|
|
|
> [!NOTE] |
|
|
> **Use "You are an assistant with reasoning capabilities." system message to consistently trigger gemini-style thinking.** |
|
|
|
|
|
> [!NOTE] |
|
|
> **I'm working on improving the dataset & model and will release a new, full version.** |
|
|
|
|
|
|
|
|
# Training Dataset |
|
|
|
|
|
- The fine-tuning dataset consists of ~450 diverse examples, 250 of which are directly from Gemini 2.5 Pro. |
|
|
|
|
|
## Trained on: |
|
|
- Unsloth version of Qwen3-30B-A3B (instruct). |
|
|
- 32k seq_len with examples ranging from 1k to ~20k tokens. |
|
|
- Up to 2 turns of conversations. |
|
|
|
|
|
--- |
|
|
|
|
|
- No benchmark data for now. |
|
|
|
|
|
**Keep in mind that it's slightly overfit since the training dataset was quite small. The model can be used to create more high quality examples for further training.** |
|
|
|
|
|
|
|
|
 |
