Training previously loaded entire wordlists into RAM and tokenized all at
once, causing OOM on large files like rockyou.txt. This adds memory
estimation, lazy dataset loading, and training optimizations.
- Add _get_available_memory_mb() for cross-platform RAM detection
- Add _estimate_training_memory_mb() to predict peak usage before loading
- Replace bulk tokenization with LazyPasswordDataset (file offset index + on-the-fly tokenization)
- Add --max-lines flag to limit training to first N lines
- Add --memory-limit flag to auto-tune --max-lines based on available RAM
- Enable gradient checkpointing and gradient accumulation (steps=4)
- Enable fp16 on CUDA devices
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add ability to fine-tune PassGPT models on custom password wordlists.
Models save locally to ~/.hate_crack/passgpt/ with no data uploaded to
HuggingFace (push_to_hub=False, HF_HUB_DISABLE_TELEMETRY=1). The
PassGPT menu now shows available models (default + local fine-tuned)
and a training option. Adds datasets to [ml] deps and passgptTrainingList
config key.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add PassGPT as attack mode 17, using a GPT-2 model trained on leaked
password datasets to generate candidate passwords. The generator pipes
candidates to hashcat via stdin, matching the existing OMEN pipe pattern.
- Add standalone generator module (python -m hate_crack.passgpt_generate)
- Add [ml] optional dependency group (torch, transformers)
- Add config keys: passgptModel, passgptMaxCandidates, passgptBatchSize
- Wire up menu entries in main.py, attacks.py, and hate_crack.py
- Auto-detect GPU (CUDA/MPS) with CPU fallback
- Add unit tests for pipe construction, handler, and ML deps check
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
