lwaekfjlk/artifact-bench

ArtifactBench

A heterogeneous graph of HuggingFace model / dataset / paper / codebase nodes (14,053) with observed (model, dataset, performance-metric) evaluation edges (51,337 relations), for benchmarking link prediction and attribute (metric-value) regression, plus an agent-based verification suite.

License

Released under the Open Database License (ODbL) v1.0 — see LICENSE or https://opendatacommons.org/licenses/odbl/1-0/. Share/modify/use freely with attribution; keep derivative databases open under ODbL.

Dataset Viewer

The viewer renders data/eval_edges.jsonl (config eval_edges): one row per (model_id, dataset_id, metric, value) over all 30,499 normalized evaluation edges. The full graph (embeddings, adjacency, splits) lives in the directories below as .npz / .json and is loaded programmatically.

Layout

path	contents
full/	unsplit graph: node_metadata.json, node_mappings.json, node_embeddings_{voyage,random}.npy (N×1024), edges{,_eval,_base_model,_resource}.npz, matching edge_metadata*.json
transductive/	edge-level split (all nodes in train & test); {train,test}_split/{edges,pos_edges}.npz, node_metadata.json, edge_metadata_normalized.json, node_embeddings_*.npy, split_info.json
inductive/	disjoint-node split (test models unseen in train); same layout + node_split.json
verification_bench/	bench.json (263 (model,dataset,metric) triples) + agent_results/<cell>/ reference agent outputs
case_study_nli/	frozen 48-model × 12-dataset NLI grid + aggregate & plotting scripts

Node types: model, dataset, paper, codebase. Edge types: model↔dataset (eval, metrics normalized to [0,1]), model↔{paper,codebase}, dataset↔{paper,codebase} (resource), model↔model (base-model/fine-tune).

Usage

from huggingface_hub import snapshot_download
import numpy as np, json
p = snapshot_download("lwaekfjlk/artifact-bench", repo_type="dataset")
emb = np.load(f"{p}/transductive/node_embeddings_voyage.npy")          # (N, 1024)
nm  = json.load(open(f"{p}/transductive/train_split/node_metadata.json"))
pos = np.load(f"{p}/transductive/train_split/pos_edges.npz")["edges"]  # (2, E)

Notes

• case_study_nli: 9 bug cells use previous_accuracy; 3 binary-output models are masked on 3-way NLI sets; 2 cells are true failures. Scripts in case_study_nli/scripts/ regenerate the aggregate + figures.
• verification_bench: 263/266 cells have a complete results.json; cell dirs named <model>_<dataset>_<metric> with /→_.