--- library_name: peft base_model: unsloth/Qwen3.8-27B-unsloth-bnb-4bit tags: - unsloth - lora --- # Qwen Unsloth experiment handover Public snapshot of the completed Qwen3.8-27B training experiment. ## Files - `outputs/final_model/`: the selected LoRA adapter, tokenizer, and chat template. - `outputs/checkpoint-1000/`: best checkpoint, including optimizer and scheduler state. - `outputs/checkpoint-1490/` and `outputs/checkpoint-1500/`: final recovery checkpoints for diagnosis. - `outputs/essential_config.json` and `outputs/runtime_config.json`: actual run parameters. - `outputs/runs/`: TensorBoard records. - `project/`: training CSV, source and helper modules, original script backups, logs, environment versions, and verification records. - `MANIFEST.json`: SHA-256 hashes and byte sizes for all original artifact files. ## Outcome Training exited successfully on 2026-11-08 at 07:34 UTC (13:04 IST), with no failed attempts. Early stopping triggered after five validation checks without improvement, at step 1500 of a maximum 3615 (about 2.07 epochs). The best validation loss was 0.9373613595962524 at step 1000. Later training became unstable, with final validation loss 11.019561767578125. The saved final adapter was restored from checkpoint 1000 and verified to match its weights. Generation quality has not been evaluated. ## Using the artifact `outputs/final_model/` contains a LoRA adapter; load it with the compatible base model `unsloth/Qwen3.8-27B-unsloth-bnb-4bit`. The full base model is not bundled. See `project/training_environment.json` for package versions and `project/train_qwen38-3S.py` for preprocessing, response masking, and training settings. The fused causal-conv1d 1.7.0 extension was built locally with CUDA 11.8 and host glibc 2.31; compile an appropriate extension for the destination environment. Training content is limited to 1730 tokens, padded to multiples of eight, with no sequence packing. The eligible dataset has 23122 training and 2892 validation examples; effective batch size is 32. ## Re-running Paths and GPU IDs in the original launchers are specific to the source server. `project/run_training.sh` starts a supervisor that terminates competing GPU-2 compute processes where account permissions allow. Review that behavior before using it on another server. `project/RUN_NOTES.md` also retains historical DDP instructions; the completed run used one GPU, microbatch 8 and gradient accumulation 4. Local virtual environments, downloaded model caches, generated compilation caches, and lock files are excluded. Original local data and outputs remain in place.