lwaekfjlk/artifact-bench

Artifact Graph

A heterogeneous graph of HuggingFace model/dataset/paper/codebase nodes with observed (model, dataset, performance-metric) evaluation edges, used to benchmark link prediction and attribute regression.

Contents

path	description
full/	Full unsplit graph: all nodes + all edges (by type)
transductive/	All nodes visible in both train and test; edges split
inductive/	Disjoint node partition: some nodes train-only, others test-only

Full graph (full/)

file	description
node_metadata.json	Per-node {type, name, downloads, info} for all 14K nodes
node_mappings.json	Integer ID ↔ HuggingFace ID mapping
node_embeddings_voyage.npy	Voyage-3 embeddings, (N, 1024)
node_embeddings_random.npy	L2-normalised random embeddings
edges.npz	All edges combined, (2, E)
edges_eval.npz	model × dataset evaluation edges
edges_base_model.npz	model → base_model edges
edges_resource.npz	model/dataset → paper/codebase edges
edge_metadata.json	Raw (model, dataset, metric) edge records
edge_metadata_normalized.json	Eval edges with metrics normalised to [0, 1]
edge_metadata_eval.json	Eval-edge metadata only
edge_metadata_base_model.json	base-model edge metadata
edge_metadata_resource.json	paper / codebase resource edge metadata

Each split directory contains:

file	description
node_embeddings_voyage.npy	Voyage-3 embeddings, shape (N, 1024)
node_embeddings_random.npy	L2-normalised random embeddings, same shape
split_info.json	Split metadata (seed, counts, dates)
node_split.json (inductive)	Per-node train/test assignment
train_split/	Training subgraph (see below)
test_split/	Test subgraph (held-out eval edges)

Each {train,test}_split/ holds:

file	description
node_metadata.json	Per-node {type, name, downloads, info}
edge_metadata_normalized.json	Normalized (u,v) → metric: value map
edges.npz	Message-passing edges, edges key, shape (2, E)
pos_edges.npz	Positive eval edges (model × dataset with metric)

Node types

• model: HuggingFace models (e.g., sileod/deberta-v3-large-tasksource-nli)
• dataset: HuggingFace datasets (e.g., nyu-mll/multi_nli)
• paper: referenced papers (arXiv IDs)
• codebase: linked repositories

Edge types

• model ↔ dataset (eval): accuracy / F1 / BLEU / etc. (normalized to [0, 1])
• model ↔ paper, model ↔ codebase, dataset ↔ paper, dataset ↔ codebase: resource links
• model ↔ model: base-model / fine-tune relations

Usage

from huggingface_hub import snapshot_download
path = snapshot_download("lwaekfjlk/artifact-graph", repo_type="dataset")
import numpy as np, json
emb = np.load(f"{path}/transductive/node_embeddings_voyage.npy")
nm = json.load(open(f"{path}/transductive/train_split/node_metadata.json"))
pe = np.load(f"{path}/transductive/train_split/pos_edges.npz")["edges"]
print(emb.shape, len(nm), pe.shape)

Case study: NLI (case_study_nli/)

Frozen 576-cell evaluation grid used as the NLI case study in our paper: 48 NLI models × 12 NLI datasets. Each cell was produced by an LLM-coder pipeline that emitted per-example predictions and a top-level accuracy.

Layout

path	description
case_study_nli/raw_evals/<model>_<dataset>_accuracy/predictions.json	Per-example {idx, prediction, ground_truth}
case_study_nli/raw_evals/<model>_<dataset>_accuracy/results.json	{accuracy: float} (plus previous_accuracy for 9 bug cells)
case_study_nli/all_results_summary_fixed.json	Cleaned aggregate: 576 rows, 9 bug fixes applied, masked flags for cells that cannot be scored 3-way
case_study_nli/scripts/rebuild_nli_summary.py	Raw → fixed aggregate
case_study_nli/scripts/plot_nli_heatmap.py	45-model heatmap (3 models with degenerate cells excluded by --min-cell 0.05)
case_study_nli/scripts/plot_nli_matrix_scree.py	Double-centered SVD scree plot
case_study_nli/figures/nli_results_heatmap.{png,pdf}	Main heatmap
case_study_nli/figures/nli_matrix_scree.{png,pdf}	Scree plot

Known issues in the raw evaluations

1. 9 bug-fix cells: top-level accuracy=0 was overwritten but previous_accuracy>0 holds the real value. The fixed aggregate uses previous_accuracy.
2. Binary-output models on 3-way datasets: three zero-shot classifiers only emit 2 labels; their MNLI / SNLI / ANLI / NLI_FEVER cells are masked in the aggregate (not directly comparable to 3-way models).
3. 2 true failures: microsoft/deberta-v3-base on allenai/scitail and araag2/MedNLI produced degenerate predictions.

Reproducibility note

The per-cell evaluation scripts were not uniformly persisted to disk — cells run in the January batch retained them, but the April re-runs (the majority) executed inline via an agent and only wrote results back. We therefore ship just the frozen outputs (predictions.json + results.json) rather than an incomplete script set. The processing scripts in case_study_nli/scripts/ are sufficient to regenerate the aggregate and figures from the per-cell outputs.

Reproduce aggregate + figures

pip install datasets numpy matplotlib huggingface_hub
python scripts/rebuild_nli_summary.py \
    --src case_study_nli/raw_evals \
    --out case_study_nli/all_results_summary_fixed.json
python scripts/plot_nli_heatmap.py \
    --input case_study_nli/all_results_summary_fixed.json \
    --out-dir case_study_nli/figures
python scripts/plot_nli_matrix_scree.py \
    --input case_study_nli/all_results_summary_fixed.json \
    --out-dir case_study_nli/figures

Verification bench (verification_bench/)

Full agent-based eval reproductions: 263 (model, dataset, metric) cells drawn from a stratified "hard" sample of the artifact graph. A skill-based multi-agent system (driver: GPT-5.2, tool mode: multiturn_metadatatool) attempts to reproduce each published accuracy score by locating the dataset, loading the model, writing an eval script, and reporting a metric.

Layout

verification_bench/
└── skills_multiagent_gpt-5.2_metadatatool/
    └── <model>_<dataset>_<metric>/
        ├── metadata.json       # (model, dataset, metric) spec
        ├── run_eval.py         # agent-written evaluation script
        ├── predictions.json    # per-example predictions
        ├── results.json        # top-level metric value
        ├── run.log             # agent trajectory log
        └── results_full.json   # rich metric breakdown (4 cells only)

Use

import json, os
ROOT = "verification_bench/skills_multiagent_gpt-5.2_metadatatool"
for cell in os.listdir(ROOT):
    meta = json.load(open(f"{ROOT}/{cell}/metadata.json"))
    result = json.load(open(f"{ROOT}/{cell}/results.json"))
    print(meta["model_id"], meta["dataset_id"], result)

Notes

• 263 / 266 cell dirs contain a complete results.json; the remaining 3 failed with agent / runtime errors.
• Cell directories are named <model>_<dataset>_<metric> with / replaced by _ in HuggingFace IDs.
• This suite is the best-performing agent configuration we evaluated (156 cells above accuracy 0.5, 97 above 0.8); scores are properly normalised to [0, 1].