mideavalwisard/llms-can-schedule

LLMs Can Schedule — Phi-3 LoRA checkpoints

LoRA adapters from the controlled Phi-3 job-shop-scheduling experiments.

The source code and frozen experiment protocol are available at:

https://github.com/starjob42/llms_can_schedule/tree/code_reproduction

The source dataset and dataset tooling are maintained at:

https://github.com/starjob42/datasetjsp

Checkpoints

Each checkpoint directory contains the LoRA adapter, tokenizer files, and inference metadata. Optimizer, scheduler, RNG, and training-state files are intentionally excluded.

• checkpoint-461: 25% of training
• checkpoint-922: 50% of training
• checkpoint-1383: 75% of training
• checkpoint-1844: 100% of training

Load an adapter by pointing PEFT at one of the checkpoint directories and the base model at microsoft/Phi-3-mini-128k-instruct. Use the exact evaluation protocol in the GitHub repository; do not interpret a checkpoint as a complete model without its base model.