smol_llama: 220M GQA
model card WIP, more details to come
A small 220M param (total) decoder model. This is the first version of the model.
- 1024 hidden size, 10 layers
- GQA (32 heads, 8 key-value), context length 2048
- train-from-scratch on one GPU :)
Links
Here are some fine-tunes we did, but there are many more possibilities out there!
- instruct
- code
- python (pypi) - link
- zephyr DPO tune
Open LLM Leaderboard Evaluation Results
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 29.44 |
| AI2 Reasoning Challenge (25-Shot) | 24.83 |
| HellaSwag (10-Shot) | 29.76 |
| MMLU (5-Shot) | 25.85 |
| TruthfulQA (0-shot) | 44.55 |
| Winogrande (5-shot) | 50.99 |
| GSM8k (5-shot) | 0.68 |
- Downloads last month
- 10
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.
Datasets used to train blockblockblock/smol_llama-220M-GQA-bpw4.2
Evaluation results
- normalized accuracy on AI2 Reasoning Challenge (25-Shot)test set Open LLM Leaderboard24.830
- normalized accuracy on HellaSwag (10-Shot)validation set Open LLM Leaderboard29.760
- accuracy on MMLU (5-Shot)test set Open LLM Leaderboard25.850
- mc2 on TruthfulQA (0-shot)validation set Open LLM Leaderboard44.550
- accuracy on Winogrande (5-shot)validation set Open LLM Leaderboard50.990
- accuracy on GSM8k (5-shot)test set Open LLM Leaderboard0.680