6uvsoomJ/blackboard-intelligence

Blackboard Intelligence

LoRA adapters accompanying Blackboard Intelligence Can Surpass Autoregressive on Globally Constrained Problems.

The adapters are for GSAI-ML/LLaDA-8B-Instruct and support three globally constrained domains:

• zebralogic/ — ZebraLogic-Hard
• nurse_rostering/ — Nurse Rostering
• jssp/ — Job-Shop Scheduling

Use with the reference code

Clone the reference implementation, then download the adapter directories into its checkpoints/ directory:

hf download 6uvsoomJ/blackboard-intelligence --include "zebralogic/*" --local-dir checkpoints
hf download 6uvsoomJ/blackboard-intelligence --include "nurse_rostering/*" --local-dir checkpoints
hf download 6uvsoomJ/blackboard-intelligence --include "jssp/*" --local-dir checkpoints

Run the domain solver from the repository root. For example:

python domains/zebralogic/solve.py \
  --checkpoint checkpoints/zebralogic \
  --eval_path data/zebralogic/zebralogic_hard_eval.json \
  --methods greedy blackboard --depth 3 --width 5

See the code repository for installation, inference, training, and evaluation details.

Files

Each domain directory contains:

<domain>/
  lora_adapter/
    adapter_model.safetensors
    adapter_config.json
  training_config.json

These are LoRA adapter weights only; the base-model files, optimizer state, and tokenizer are not included.

License

The adapters in this repository are released under the MIT License. Use of GSAI-ML/LLaDA-8B-Instruct is additionally subject to the base model's license and terms.

Citation

Citation information will be added upon arXiv publication.