gnn-ruby-code-study / specs /gnn_topology.yaml
timlawrenz's picture
Upload specs/gnn_topology.yaml with huggingface_hub
71de54d verified
# Research Spec: GNN Decoder Topology — Track 4
#
# Tests tree-aware decoding: giving the decoder the actual AST edge structure
# instead of sequential chain edges. Three modes:
# - chain: Legacy baseline (0→1→2→…)
# - teacher_forced: Ground-truth AST edges during GNN message passing
# - iterative: Two-pass: chain→predict parents→rebuild tree→refine
#
# The hypothesis: the decoder's GNN never sees tree topology, so it cannot
# learn structure-sensitive generation. Providing real edges should improve
# parent prediction accuracy and node-type diversity.
#
# Launch:
# ratiocinator research specs/gnn_topology.yaml
# What to research
topic: "Decoder topology for GNN code generation: does giving the decoder the real AST tree structure (instead of sequential chain edges) improve reconstruction? Compare chain baseline, teacher-forced ground-truth edges, and iterative predict-then-refine. Cross with GAT/GCN/GIN decoder architectures and improved/comprehensive loss functions."
goal_metric: syntactic_validity_pct
maximize: true
# Target codebase
repo_url: https://github.com/timlawrenz/jubilant-palm-tree.git
repo_branch: experiment/ratiocinator-gnn-study
runner_script: scripts/run_topology_arm.sh
# Infrastructure — ~850K params, moderate training
hardware:
gpu: "RTX 4090"
num_gpus: 1
min_cpu_ram_gb: 32
min_inet_down: 1000.0
min_cuda_version: 12.0
max_dph: 0.40
disk_gb: 50.0
image: pytorch/pytorch:2.7.0-cuda12.8-cudnn9-runtime
data:
source: none # Dataset is in the repo branch
deps:
pre_install:
- "apt-get update -qq && apt-get install -y -qq git-lfs > /dev/null 2>&1 || true"
- "cd /workspace/experiment && git lfs install && git lfs pull"
- "pip install torch-geometric torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-2.7.0+cu128.html"
- "pip install pandas tqdm sentence-transformers nltk scikit-learn numpy"
requirements: requirements.txt
exclude_from_requirements:
- torch
- torchvision
- torch_geometric
verify: "python -c \"import torch_geometric; print(f'PyG {torch_geometric.__version__}')\""
metrics:
protocol: json_line
json_prefix: "METRICS:"
# Budget — ~9 arms (3 edge modes × 3 conv types), ~10 min each
max_iterations: 2
max_dollars: 15.00
train_timeout_s: 2400
download_timeout_s: 600
# Output
paper_title: "What Graph Neural Networks Can and Cannot Learn About Code: A Systematic Empirical Study on Ruby AST Analysis"