qwen3-0.6b-contextual-chunks-checkpoints

Training checkpoints for p4ulbr4dl3y/qwen3-0.6b-contextual-chunks.

Fine-tuning ran on Colab, where spot instances are killed without warning, so every checkpoint was pushed here to keep the run resumable.

The mirror holds 15 checkpoints at steps 160 through 676, about 74 MB each. Alongside the adapter weights each one carries the optimizer state, the scheduler state, the RNG state and the gradient scaler, which is everything needed to resume training exactly where it stopped. logs/train_log.jsonl holds the run log.

The best checkpoint by evaluation loss is step 640, and its adapter is what the public model repository carries. The repository is kept as a record of the run; the optimizer state it stores is only useful for resuming training.

Links

• Technical report and retrieval harness: https://github.com/p4ulbr4dl3y/contextual-chunks
• Model: https://huggingface.co/p4ulbr4dl3y/qwen3-0.6b-contextual-chunks
• Training data: https://huggingface.co/datasets/p4ulbr4dl3y/ru-contextual-chunks