Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Update bamba.md - Safety results #2

Merged
merged 1 commit into from
Dec 17, 2024
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 13 additions & 10 deletions bamba.md
Original file line number Diff line number Diff line change
Expand Up @@ -100,16 +100,19 @@ We compare Bamba and Falcon Mamba with SoTA transformer models of similar size (

### Safety tasks

Safety benchmarks are crucial for ensuring AI models generate content that is ethical, inclusive, and non-harmful. We evaluate our model on well known safety benchmarks such as Toxigen (5-shot, logits) (focused on detecting toxic language), BBQ (5-shot, generation), PopQA (5-shot, generation), and Ethos (which measures bias and fairness). These benchmarks help us identify and mitigate harmful outputs, ensuring the model avoids generating offensive or discriminatory content. We intend to fix the gaps in safety through comprehensive SFT and DPO approaches.

| Model | PopQA | Toxigen | BBQ |
| :---- | :---- | :---- | :---- |
| [Bamba 9B](https://huggingface.co/ibm-fms/Bamba-9B) | 20.5 | 57.4 | 44.2 |
| [Olmo2 7B](https://huggingface.co/allenai/OLMo-2-1124-7B) | 25.7 | 63.1 | 58.4 |
| [Gemma2 9B](https://huggingface.co/google/gemma-2-9b) | 27.3 | 69.6 | 59.9 |
| [IBM Granite v3 8B](https://huggingface.co/ibm-granite/granite-3.0-8b-base) | 27.5 | 79.9 | 82.1 |
| [Meta Llama 3.1 8B](https://huggingface.co/meta-llama/Llama-3.1-8B) | 28.8 | 67 | 60 |
| [Falcon Mamba 7B](https://huggingface.co/tiiuae/falcon-mamba-7b) | 19.3 | 62.1 | 60.2 |
Safety benchmarks are crucial for ensuring AI models generate content that is ethical, inclusive, and non-harmful. We evaluate our model on well known safety benchmarks such as Toxigen (5-shot, logits) (focused on detecting toxic language), BBQ (5-shot, generation), PopQA (5-shot, generation), and CrowS-Pairs (which measures bias and fairness). We intend to fix the gaps in safety through comprehensive SFT and DPO approaches.

| Model | PopQA | Toxigen | BBQ | Crow-SPairs* |

| :---- | :---- | :---- | :---- | :---- |
| [Bamba 9B](https://huggingface.co/ibm-fms/Bamba-9B) | 20.5 | 57.4 | 44.2 | 70.7 |
| [Olmo2 7B](https://huggingface.co/allenai/OLMo-2-1124-7B) | 25.7 | 63.1 | 58.4 | 72 |
| [Gemma2 9B](https://huggingface.co/google/gemma-2-9b) | 27.3 | 69.6 | 59.9 | 71.7 |
| [IBM Granite v3 8B](https://huggingface.co/ibm-granite/granite-3.0-8b-base) | 27.5 | 79.9 | 82.1 | 75 |
| [Meta Llama 3.1 8B](https://huggingface.co/meta-llama/Llama-3.1-8B) | 28.8 | 67 | 60 | 70.8 |
| [Falcon Mamba 7B](https://huggingface.co/tiiuae/falcon-mamba-7b) | 19.3 | 62.1 | 60.2 | 75 |

*Lower is better

## Comparison with transformers with similar token budget

Expand Down