innodata-llm-safety

KO | EN

gitlite — search

#javascript #typescript #rust #linux #docker #macos #nodejs #java #windows #awesome #pytorch #mcp #ios #openai

innodata-llm-safety

(★ 10)

Benchmarking Llama, Mistral, Gemma and GPT for Factuality, Toxicity, Bias and Propensity for Hallucinations

innodata-llm-safety Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# innodata-llm-safety
Benchmarking Llama, Mistral, Gemma, DeepSeek and GPT for factuality, toxicity, bias, instruction following, avoiding jailbreaks and propensity for hallucinations

| [LLM Safety Benchmark](https://llm-safety.innodata.com/) |  [Paper](https://arxiv.org/abs/2404.09785) | [39 Safety Datasets](https://huggingface.co/innodatalabs)  |  [Red teaming OSS Tool](https://github.com/innodatalabs/redlite) | 
|---|---|---|---|
|  <a href="https://llm-safety.innodata.com/"><img style="border:1px solid black;" src="img/results.png" alt="drawing" width="300"/></a> |  <a href="https://arxiv.org/abs/2404.09785"><img style="border:1px solid black;" src="img/paper.png" alt="drawing" width="300"/></a> | <a href="https://huggingface.co/innodatalabs"><img style="border:1px solid black;" src="img/datasets.png" alt="drawing" width="300"/></a>  | <a href="https://github.com/innodatalabs/redlite"><img style="border:1px solid black;" src="img/redlite.png" alt="drawing" width="300"/></a>  |

> [!NOTE]
>**UPDATED** May 5th, 2026:
> - Comparing latest major models Gemini 3.1 pro, Claude Opus 4.7, GPT 5.4
> - Comparing latest small models  Hunyuan, Llama, Phi, Gemma, Olmo, Mistral, gpt-oss
> - 10 new state-of-the-art datasets for agentic, code and function calling (rt5 series)

> **UPDATED** January 7th, 2026:
> - Comparing latest major models Gemini 3 pro, Claude Opus 4.5, GPT 5.1, Deepseek 3.1 
> - Comparing latest small models Qwen, Hunyuan, Llama, Phi, Gemma, Olmo, Mistral, gpt-oss 
> - 4 new state-of-the-art datasets for stem, code and function calling (rt4 series)

> **UPDATED** August 18th, 2025:
> - Comparing latest Qwen, DeepSeek, Llama, Phi, Gemma, Olmo, Mistral 
> - Benchmarking OpenAI open source model: gpt-oss-20b 
> - 3 new state-of-the-art safety datasets (rt3 series)

> **UPDATED** March 12th, 2025:
> - Added Qwen2.5-7B

> **UPDATED** February 10th, 2025:
> - Benchmarking latest open source LLMs including *DeepSeek*, OLMo-2, etc.
> - All datasets revised: 100s of ground truth corrections.
> - Model tested are all 'small LLMs' (7B-12B parameters) except for GPT-4o added to the benchmark as 'upper limit'.

> **UPDATED** August 19th, 2024:
> - Benchmarking latest open source LLMs: Gemma-2, Llama3 & 3.1, Mistral v0.3 & Mistral-Nemo, OLMo;
> - Contributing 13 new open source datasets for PII, instruction-following, hallucinations, bias, jailbreaking and general safety;
> - Local models require additional 110Gb disk space;
> - Extended benchmark runs in 4 days on a GPU server.

# Reproducing our Research

## Required hardware

We ran the benchmark on a server with 1 x NVIDIA A100 80GB.

Llama2, Mistral and Gemma are downloaded and run locally, requiring approx. 90Gb disk space.

## Set up

```bash
python3.11 -m venv .venv
. .venv/bin/activate
pip install wheel pip -U
pip install -r requirements.txt
```

(Works on Python3.10 as well.)

## Required tokens and environment variables

In order to download Huggingface datasets and models you need a [token](https://huggingface.co/settings/tokens).

Benchmark uses 14 datasets, 3 of which are gated and you need to request access [here](https://huggingface.co/datasets/innodatalabs/rt-inod-finance), [here](https://huggingface.co/datasets/innodatalabs/rt-inod-bias) and [here](https://huggingface.co/datasets/innodatalabs/rt-inod-jailbreaking).

Llama2 is gated model, you need to [request access](https://llama.meta.com/llama-downloads/).

Gemma is gated model, you need to [request access](https://www.kaggle.com/models/google/gemma/license/consent).

In order to call OpenAI API, you need a [key](https://platform.openai.com/api-keys).

Export secret keys to environment variables:

```bash
export HF_TOKEN=xyz
export OPENAI_API_KEY=xyz
```

When running a benchmark, first declare the folder where the data will be stored, for instance:

```bash
export REDLITE_DATA_DIR=./data
```

## Run the benchmark

The following script does it all:

```bash
python run_all.py
```

Original benchmark runs in ~24 hours on a GPU server.

## Visualize

Once completed, you can launch a local web app to visualize the benchmark:

```bash
redlite server
```

innodata-llm-safety

innodata-llm-safety Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

innodata-llm-safety

innodata-llm-safety Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README