Problem
Classical optimization problems carry hard feasibility constraints that vanilla language models often violate.
NLP & ML Research · 202570h → 11.5h fine-tuning
LLMs for Job-Shop Scheduling
Course research on fine-tuning Llama 3.1 8B with graph-augmented prompts for job-shop scheduling (JSSP).
Approach
Our team of three modeled JSSP instances as disjunctive graphs, serialized precedence and machine conflicts into the prompt, and fine-tuned Llama 3.1 8B with 4-bit QLoRA.
Result
The pipeline produced a much more structured optimization workflow and made training/runtime tradeoffs manageable for course-scale experimentation.
Step 1 of 5: JSSP instance
Highlights
- Fine-tuned Llama 3.1 8B with 4-bit QLoRA (Unsloth) on an NVIDIA RTX A6000, in a graph-augmented pipeline that converts each instance into a disjunctive graph (PyTorch Geometric) serialized into the prompt.
- Pre-processed the noisy STARJOB dataset (~130k to 9,525 usable instances), cutting fine-tuning runtime from 70h to 11.5h.
Course research project under Prof. Aijun An at York University, built in a team of three.
What I built
I treated scheduling as a structure-first problem instead of a pure prompting problem. The project converts job shop instances into graph-shaped representations so the model sees precedence and machine conflicts explicitly.
Core decisions
- Serialized machine and operation constraints instead of relying on free-form descriptions.
- Used disjunctive graphs to preserve conflict structure.
- Applied 4-bit QLoRA fine-tuning to keep experimentation feasible.
- Focused on failure slices where outputs looked plausible but violated constraints.
What mattered most
The project became stronger once I stopped asking whether the LLM could "solve optimization" in the abstract and started asking whether the representation exposed enough constraint information for the model to reason over.