IMJOHNADAMS/discovery-agent-eval-trajectories

Discovery Agent Evaluation Trajectories

This dataset contains 1,728 recorded discovery-agent trajectories across 6 model identifiers, 24 tasks, and 4 harness labels. The original JSON files are preserved byte for byte, with their original directory structure and field values.

Files and data loading

• trajectories/<model>/<task>/<harness>__run<N>.json: complete original records, including all recorded steps, decisions, generations, evaluation results, best code, metadata, and belief-state history.
• data/trajectories.parquet: a compact index with one row per trajectory.

The Dataset Viewer and the automatically generated Croissant metadata describe the index. Each row links to its complete original JSON using raw_path and raw_url, and includes its byte size and SHA-256 checksum. Nested step contents are available in those JSON files; they are not expanded into Croissant fields. The train split is a conventional container for the complete inventory, not an experimental train/test partition or a recommendation to use the data for training.

import json
from datasets import load_dataset
from huggingface_hub import hf_hub_download

repo_id = 'IMJOHNADAMS/discovery-agent-eval-trajectories'
index = load_dataset(repo_id, split="train")
row = index[0]
path = hf_hub_download(repo_id, row["raw_path"], repo_type="dataset")
with open(path, encoding="utf-8") as f:
    trajectory = json.load(f)

Some original files are large (the largest is 1,907,494,602 bytes). Use a streaming JSON reader when processing these files with limited memory.

Index fields

Field	Meaning
trajectory_id	Unique source-relative path without the .json extension.
model	Original JSON model identifier.
task	Task directory name.
task_id	Original JSON task identifier, which can differ from the directory name.
harness	Harness label from the filename, preserved as recorded.
level	Original JSON level value, which can differ from the filename label.
run	Run number from the filename.
run_id	Original JSON run identifier.
num_steps	Number of recorded steps.
step_budget	Original step budget.
best_score	Original best score; score scales depend on the task.
tokens_used	Original total token accounting value.
total_prompt_tokens	Original prompt token accounting value.
total_completion_tokens	Original completion token accounting value.
started_at, finished_at	Original timestamps; missing values remain null.
raw_path	Repository-relative path of the complete original JSON.
raw_url	Download URL of that JSON on the repository's main branch.
raw_size_bytes	Original file size in bytes.
raw_sha256	SHA-256 checksum of the original file bytes.

Coverage

Total original JSON bytes: 8,002,924,353. Total recorded steps: 251,368.

Model	Trajectories
claude-opus-4-6	288
deepseek-v3.2	288
gemini-3.1-pro-preview	288
gpt-5.4	288
gpt-oss-120b	288
qwen3-next-80b-instruct	288

Harness filename labels: L0 (432), L0f (432), L1 (432), L2b (432). These labels and the JSON level values have not been renamed.

Croissant metadata

The repository also includes croissant.json, generated and validated with the MLCommons mlcroissant library against the published Parquet index. Its distribution URL is pinned to an immutable repository commit and its SHA-256 checksum is included. It describes all 1,728 index records and 20 scalar fields, with links to the complete original JSON files.

After the Dataset Viewer has processed the index, download the automatically generated Croissant JSON-LD document from:

https://huggingface.co/api/datasets/IMJOHNADAMS/discovery-agent-eval-trajectories/croissant