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	Paper	39 Safety Datasets	Red teaming OSS Tool

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

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.

Benchmark uses 14 datasets, 3 of which are gated and you need to request access here, here and here.

Llama2 is gated model, you need to request access.

Gemma is gated model, you need to request access.

In order to call OpenAI API, you need a key.

Export secret keys to environment variables:

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:

export REDLITE_DATA_DIR=./data

Run the benchmark

The following script does it all:

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:

redlite server