⚠️ ILLEGAL OR UNETHICAL USE IS PROHIBITED!
Sep 10, 2024
Salut Lynx!
Lynx is a set of fast and advanced conversational models for cybersecurity professionals fine-tuned from the incredible Mistral 7B and Mistral 8x7B (Coming Soon).
Open Now. Open Forever.
I strongly believe in the potential of open source to drive innovation and foster collaboration in the field of AI. As a result, I am releasing Petit Lynx, and the majority of the datasets used in training under the Apache 2.0 license, enabling everyone to utilize the model and datasets without any limitations under ethical use.
Petit Lynx & Grande Lynx
Lynx comprises two models, specifically fine-tuned versions of Mistral-7B and Mistral-8x7B, which are trained on synthetically generated datasets. These adaptations equip the models with insights from historical cybersecurity incidents including breaches and attacks, notable personalities in the cybersecurity arena, programming vulnerabilities, mathematical word problems, and technology-related incident management.
Cyber Logic
Marmottes Données
Code Malade
Cyber Logic Marmottes Données Code Malade
Fine-tuned on over 300,000+ records. All training data is free, public, and published open source on HuggingFace under the Apache 2 license.
Model Evaluations
Context Length
32K
Train Loss
0.65
Perplexity
1.67
Eval Loss
0.74
Training Parameters
| Hyperparameter | Value |
|---|---|
| learning_rate | 5.e-6 |
| weight_decay | 0.01 |
| pct_start | 0.1 |
| seq_len | 8000 |
| batch_size | 1 |
| max_steps | 25500 |
| seed | 44 |
| distributed_type | multi-GPU |
| num_devices | 4 |
| num_epochs | 7.3 |