AI Generated Notes:
What's Changed
- Fix --scale_wd inv_linear being silently ignored by @TonyChen06 in #51
- Save evaluator plots for every seed, not just the last one by @TonyChen06 in #54
- Count LR schedule horizon in optimizer steps, not micro-batches by @TonyChen06 in #55
- Add HFLLM base class; make per-family LLM wrappers thin subclasses by @TonyChen06 in #57
- Cast connector inputs to the live weight dtype by @TonyChen06 in #61
- Surface crash errors in wandb instead of masking as success by @willxxy in #65
- Fix silent ECG drop in chat by sharing the signal-injection ELM set (Issue #66) by @willxxy in #67
- Add SigLIP-1D ECG encoder receiver (--encoder siglip-ecg) by @TonyChen06 in #69
- Add Qwen3 support (0.6B/1.7B/4B/8B) by @TonyChen06 in #70
- Feat/qwen3.5 support by @TonyChen06 in #71
Full Changelog: 0.0.0...0.0.1
Human Notes:
- Many bug fixes as noted above.
- Added latest Qwen models and adapted SigLip2 to natively process ECG.
- Made RL training more stable.
Contributors
willxxy and TonyChen06