romaingrx/red-teamer-mistral-nemo

This model is a fine-tuned version of mistralai/Mistral-Nemo-Instruct-2407 on the wildjailbreak dataset. Only the adversarial_harmful data types have been used for training.

Uses

This model is intended to be used for red-teaming purposes only. It generates prompts that are likely to evade existing LLMs' content filters based on the user's input.

→ The HarmBench evaluation will be released soon.

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 16
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 1

Results

Epoch	Step	Validation Loss	Training Loss
0.0982	20	1.2933	1.3425
0.1965	40	1.1966	1.2067
0.2947	60	1.1594	1.1544
0.3930	80	1.1386	1.1427
0.4912	100	1.1259	1.1235
0.5895	120	1.1179	1.1167
0.6877	140	1.1129	1.1153
0.7860	160	1.1098	1.1118
0.8842	180	1.1086	1.1112
0.9825	200	1.1083	1.1113