60 lines
2.3 KiB
Markdown
60 lines
2.3 KiB
Markdown
---
|
|
base_model: Ma7ee7/Meet7_0.6b
|
|
tags:
|
|
- text-generation-inference
|
|
- transformers
|
|
- unsloth
|
|
- qwen3
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
---
|
|
|
|
# Meet7 0.6B — Experimental
|
|
|
|
A continued fine-tune of [Meet7 0.6B](https://huggingface.co/Ma7ee7/Meet7_0.6b), trained at a lower learning rate on the same 600-sample dataset. Trades Meet7's sharp BoolQ spike for more balanced commonsense and reasoning gains across the board.
|
|
|
|
## Benchmarks
|
|
|
|
<img src="https://cdn-uploads.huggingface.co/production/uploads/6466047a326128fd2c693cfa/KfI9qNkT6jPkuquBL39UT.png" width="600"/>
|
|
|
|
0-shot evaluation, scores are `acc_norm`.
|
|
|
|
| Task | Qwen3-0.6B (Base) | Meet7 0.6B | Experimental | Δ vs Base |
|
|
|------|:-----------------:|:----------:|:------------:|:---------:|
|
|
| BoolQ | 0.3798 | **0.5554** | 0.3991 | +01.93% |
|
|
| ARC Easy | 0.3384 | 0.3952 | **0.3965** | +05.81% |
|
|
| ARC Challenge | 0.2841 | **0.3285** | 0.3259 | +04.18% |
|
|
| HellaSwag | 0.3981 | 0.4205 | **0.4265** | +02.84% |
|
|
| PIQA | 0.6338 | 0.6583 | **0.6687** | +03.49% |
|
|
| Winogrande | 0.5225 | 0.5201 | **0.5304** | +00.79% |
|
|
|
|
<details>
|
|
<summary>What these measure</summary>
|
|
|
|
- **BoolQ** — Reading comprehension and yes/no factual grounding
|
|
- **ARC Easy / Challenge** — Grade-school science reasoning; Challenge is the retrieval-resistant subset
|
|
- **HellaSwag** — Commonsense sentence completion
|
|
- **PIQA** — Physical world intuition
|
|
- **Winogrande** — Commonsense pronoun resolution
|
|
|
|
</details>
|
|
|
|
## vs Meet7 0.6B
|
|
|
|
This model is more **balanced** than Meet7. It outperforms Meet7 on HellaSwag, PIQA, and Winogrande — the physical and commonsense intuition tasks — at the cost of Meet7's large BoolQ advantage. If you need consistent commonsense reasoning, prefer this model. If yes/no QA is your primary use case, prefer Meet7.
|
|
|
|
## Model Details
|
|
|
|
| | |
|
|
|---|---|
|
|
| **Developed by** | Ma7ee7 |
|
|
| **License** | Apache-2.0 |
|
|
| **Base model** | Ma7ee7/Meet7_0.6b |
|
|
| **Original base** | unsloth/Qwen3-0.6B-unsloth-bnb-4bit |
|
|
| **Training samples** | 600 |
|
|
| **Training** | Continued LoRA fine-tune, lower LR |
|
|
|
|
Trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Hugging Face TRL.
|
|
|
|
[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth) |