experiments/gnn_decoder_topology.yaml

# Fleet Spec: Decoder Topology Fix (Track 4)
#
# Tests tree-aware decoding vs chain baseline across architectures and loss functions.
# The critical insight: decoder uses chain edges (0→1→2→...) instead of tree edges.
#
# Launch:
#   ratiocinator fleet run experiments/gnn_decoder_topology.yaml

name: gnn-decoder-topology
description: "Tree-aware decoder topology vs chain baseline"

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

repo:
  url: https://github.com/timlawrenz/jubilant-palm-tree.git
  branch: experiment/ratiocinator-gnn-study
  clone_depth: 1

data:
  source: none

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__}')\""

arms:
  # ── Edge mode comparison (GAT decoder, improved loss) ──
  - name: chain-gat
    description: "Chain edges (legacy baseline), GAT decoder"
    command: "bash scripts/run_topology_arm.sh"
    env:
      DECODER_EDGE_MODE: "chain"
      DECODER_CONV_TYPE: "GAT"
      LOSS_FN: "improved"
      HIDDEN_DIM: "256"
      NUM_LAYERS: "5"
      EPOCHS: "30"

  - name: teacher-forced-gat
    description: "Teacher-forced tree edges, GAT decoder"
    command: "bash scripts/run_topology_arm.sh"
    env:
      DECODER_EDGE_MODE: "teacher_forced"
      DECODER_CONV_TYPE: "GAT"
      LOSS_FN: "improved"
      HIDDEN_DIM: "256"
      NUM_LAYERS: "5"
      EPOCHS: "30"

  - name: iterative-gat
    description: "Iterative predict→refine, GAT decoder"
    command: "bash scripts/run_topology_arm.sh"
    env:
      DECODER_EDGE_MODE: "iterative"
      DECODER_CONV_TYPE: "GAT"
      LOSS_FN: "improved"
      HIDDEN_DIM: "256"
      NUM_LAYERS: "5"
      EPOCHS: "30"

  # ── Best edge mode × different architectures ──
  - name: teacher-forced-sage
    description: "Teacher-forced tree edges, SAGE decoder"
    command: "bash scripts/run_topology_arm.sh"
    env:
      DECODER_EDGE_MODE: "teacher_forced"
      DECODER_CONV_TYPE: "SAGE"
      LOSS_FN: "improved"
      HIDDEN_DIM: "256"
      NUM_LAYERS: "5"
      EPOCHS: "30"

  - name: teacher-forced-gin
    description: "Teacher-forced tree edges, GIN decoder"
    command: "bash scripts/run_topology_arm.sh"
    env:
      DECODER_EDGE_MODE: "teacher_forced"
      DECODER_CONV_TYPE: "GIN"
      LOSS_FN: "improved"
      HIDDEN_DIM: "256"
      NUM_LAYERS: "5"
      EPOCHS: "30"

  # ── Loss function × topology interaction ──
  - name: teacher-forced-gat-comprehensive
    description: "Teacher-forced, GAT, comprehensive loss"
    command: "bash scripts/run_topology_arm.sh"
    env:
      DECODER_EDGE_MODE: "teacher_forced"
      DECODER_CONV_TYPE: "GAT"
      LOSS_FN: "comprehensive"
      HIDDEN_DIM: "256"
      NUM_LAYERS: "5"
      EPOCHS: "30"

metrics:
  protocol: json_line
  json_prefix: "METRICS:"

budget:
  max_dollars: 10.00
  train_timeout_s: 3600
  download_timeout_s: 600