graphjepa-psf-requests-200

Precomputed (TemporalGraph, node_features) cache for the graphjepa project. Avoids repeated git-checkout + tree-sitter + BERT-feature preprocessing on every training run. Drop into graphjepa/cache/graph_<basename>_n<N>_ast.pkl and training loops will auto-hit the cache with no model downloads required.

Contents

File	Description
graph_psf_requests_n200_ast.pkl	Pickle of {'graph': TemporalGraph, 'features': dict_by_kind}

Source

• Source repo: ./data/psf_requests
• First n_commits: 200
• Source HEAD at build time: 514c1623fefff760bfa15a693aa38e474aba8560
• AST expansion: enabled
• Features: BERT base uncased, frozen embedding lookup (not forward pass)
  • content_vec: BERT-mean-pool of node.content
  • type_vec: BERT-mean-pool of node.type_description

SHA-256

2cca40c0bf0621005b90f96c091f73c4d15519eeed3a88896b5e6fc7619f087d

Usage

import os, pickle
from huggingface_hub import hf_hub_download

os.makedirs('graphjepa/cache', exist_ok=True)
path = hf_hub_download(
    repo_id="IDMedicine/graphjepa-psf-requests-200",
    filename="graph_psf_requests_n200_ast.pkl",
    repo_type="model",
    local_dir='graphjepa/cache',
)

# Once in graphjepa/cache, load_or_build_graph() auto-hits:
from graphjepa.graph_cache import load_or_build_graph
graph, features = load_or_build_graph(
    './data/psf_requests', n_commits=200, expand_ast=True)

Requires the graphjepa package to be importable so the pickled dataclasses resolve. Install from the source tree:

pip install -e /path/to/code-transformer