psssd-cat/HaloGuard-multitask-v1

HaloGuard Multitask v1

Multi-head safety classifier trained from Qwen/Qwen3.5-2B using the multitask guard trainer.

Training data

• Preprocessed dataset: ./data
• Format: multitask_structured
• Categories loaded from category_vocab.json

Training config

• Epochs: 1
• Batch size: 512
• Gradient accumulation: 1
• Effective batch: 512
• Backbone LR: 5e-5
• Head LR: 5e-4
• Scheduler: cosine
• Warmup ratio: 0.05
• Weight decay: 0.01
• Max grad norm: 1.0
• Max length: 512
• Dropout: 0.1
• Lambda cat: 0.1
• Gradient checkpointing: true
• Packed sequences: true
• LoRA: false

Outputs

• model.pt
• multitask_config.json
• category_vocab.json
• multitask_train_metadata.json

Notes

Training completed successfully. The original run crashed after training during metadata writing due to a post-run bug in train_halo_guard_multitask.py; model weights were still saved.