felixwangg/Qwen2.5-Coder-7B-contrastive-visualize-diffmask-t03

Contrastive Pre-trained Model

Base model weights after contrastive pre-training. The projection head has been removed — these weights can be used directly as base_model in downstream SFT configs.

Training Configuration

{
  "base_model": "Qwen/Qwen2.5-Coder-7B-Instruct",
  "secure_dataset": "felixwangg/prime_vul_plus_splitted",
  "insecure_dataset": "felixwangg/prime_vul_minus_splitted",
  "loss_type": "supcon",
  "temperature": 0.3,
  "pooling_strategy": "last_token",
  "projection_dim": 256,
  "embedding_layer": -1,
  "learning_rate": 8e-05,
  "num_train_epochs": 5,
  "contrastive_weight": 1.0,
  "lm_loss_weight": 0.0,
  "use_lora": true,
  "lora_r": 32,
  "lora_alpha": 32,
  "trainable_modules": null
}