hybrid50 / README.md
anonymous-quantinv's picture
Publish Hybrid50 contextual dataset with provenance and validation
58f8c8a verified
|
Raw History Blame Contribute Delete
8.21 kB
---
pretty_name: Hybrid50
license: other
license_name: hybrid50-component-licenses
license_link: LICENSE
language:
- en
task_categories:
- question-answering
size_categories:
- n<1K
tags:
- retrieval-augmented-generation
- tool-use
- agent-evaluation
- multi-agent
- prompt-injection
configs:
- config_name: default
data_files:
- split: test
path: data/test.parquet
---
# Hybrid50
Hybrid50 contains 50 English tasks that combine a knowledge question with actions
in a simulated AgentDojo environment. The knowledge answer selects a business
record or supplies the content of an action, such as sending an email, updating
a file, booking a hotel, or scheduling a payment. Four tasks use the answer as
the content of a fact-checking note rather than selecting a candidate record.
This release freezes `hybridqa-agentdojo/contextual`. It includes 20 workspace,
14 travel, and 16 banking tasks; 43 require writes and seven are read-only.
Every task includes 12 retrieval documents (600 total), its initial environment,
a scoring contract, and a reference solution. Business fixtures are synthetic.
Stable IDs run from `hybrid-001` through `hybrid-050`.
## Load the dataset
```bash
pip install datasets
```
```python
import json
from datasets import load_dataset
ds = load_dataset("anonymous-quantinv/hybrid50", split="test")
row = ds[0]
task = {
"id": row["id"],
"schema_version": row["schema_version"],
"query": row["query"],
**{field: json.loads(row[field + "_json"]) for field in (
"knowledge", "workspace", "evaluation_contract", "reference_solution"
)},
}
print(task["query"])
```
The Parquet view uses JSON strings for heterogeneous nested objects so that all
three environment schemas load reliably with Hugging Face Datasets. Decoding
the four `*_json` fields reconstructs the complete canonical record exactly.
Convenience columns expose `scenario`, `source_task_id`, `question`,
`reference_answer`, and `evidence_status` for browsing.
The canonical [tasks.jsonl](tasks.jsonl) is an unchanged copy of the frozen
dataset and can also be read with only the Python standard library:
```python
import json
from pathlib import Path
tasks = [json.loads(line) for line in Path("tasks.jsonl").read_text().splitlines()]
assert len(tasks) == 50
```
## Canonical record schema
| Field | Contents |
| --- | --- |
| `id` | Stable task ID. |
| `schema_version` | `hybridqa-agentdojo/contextual`. |
| `query` | User-facing work request including the knowledge question. |
| `knowledge` | Question, reference answer, aliases, reference period, source task ID, and complete retrieval corpus. |
| `workspace` | Scenario, AgentDojo source descriptor, source environment ID, and initial state. |
| `evaluation_contract` | Required effects, allowed reads/writes, answer-to-record binding, and report requirements. |
| `reference_solution` | Reference tool-call sequence and final response. |
Give the agent the `query`, access to retrieval over `knowledge.rag_corpus`, and
tools backed by `workspace.initial_state`. Keep reference answers, aliases,
contracts, source-review labels, and reference solutions on the evaluator side.
Do not send a complete dataset row to the agent.
## Protocol and scoring
The original retrieval configuration returns top four documents, delivers at
most 858 characters per document, and allows one retrieval call per agent per
phase. Documents retain the prepared SealQA excerpts without additional
rewriting or supplementation. AgentDojo package version `0.1.35`, benchmark
version `v1.2.2`, supplies the simulated tools and environment behavior.
Task success checks the knowledge answer, requested final report, successful
write calls when required, and their effects in the final environment. Safety
is scored separately against the allowed reads and writes. A task can succeed
while producing an unauthorized side effect. Semantically equivalent answers
that the deterministic scorer misses may require documented human review.
This is a **dataset release**: it includes the contracts and reference plans,
but does not bundle the MADBench runner, its Python adapters, or an executable
standalone scoring engine. The 50 tasks form one evaluation split; no training
or validation split is defined.
## Evidence quality and validation
The supplied excerpt review labels 36 tasks `supported`, nine `limited`, and
five `missing`. Thus 14 tasks have material evidence limitations. The five
`missing` tasks are `hybrid-008`, `hybrid-038`, `hybrid-042`, `hybrid-047`, and
`hybrid-048`. Time-sensitive answers retain their original reference periods;
they are not claims about current events at the time of use.
See [the per-task evidence review](docs/source_review.zh.md) and
[machine-readable audit](audit/revision_audit.json). The review was assisted
by an AI assistant; it is not an independent human annotation study.
Before publication, offline reference replay passed for all 50 tasks, including
256 negative checks, 286 positive checks, and 143 safety checks. The existing
35 dataset, binding, attack, and integration tests also passed. These checks
establish data consistency and reference-plan executability. They do not
measure model performance or prove that all answers can be recovered from the
provided excerpts. See [the replay report](audit/verification.json).
## Included materials
- [tasks.jsonl](tasks.jsonl): canonical nested records.
- [data/test.parquet](data/test.parquet): lossless view for the Hub viewer and Datasets library.
- [docs/tasks.zh.md](docs/tasks.zh.md): task catalog, contracts, and reference responses.
- [attacks/attack_plan.json](attacks/attack_plan.json): 800 reviewed attack entries, covering 50 tasks × four mechanisms (M2/M3/M4/M6) × four goals (G1–G4).
- [attacks/payloads.json](attacks/payloads.json) and [docs/attacks.zh.md](docs/attacks.zh.md): task-specific attack materials. These are experimental inputs, not instructions to the dataset reader or measured attack outcomes.
- [manifest.json](manifest.json): schema, provenance, counts, and file checksums.
- [scripts/verify_release.py](scripts/verify_release.py): local integrity verification.
- [scripts/upload.py](scripts/upload.py): repeatable upload using the locally authenticated Hugging Face account.
Historical supplements, superseded tasks, run logs, model credentials, and
experimental result archives are excluded from this release.
## Sources and licenses
Knowledge records reuse 50 distinct entries from the source project's prepared
100-question SealQA pool. Their question, reference answer, reference period,
and entire retrieval corpus are unchanged. The upstream
[SEALQA dataset](https://huggingface.co/datasets/vtllms/sealqa) is pinned to
`b2ecadf036972d8471a5c4cdf92aa3b3c6ba96e7`; prepared excerpts originate from its
`longseal` data. SEALQA is distributed under Apache-2.0, with its license retained
in [licenses/SEALQA-APACHE-2.0.txt](licenses/SEALQA-APACHE-2.0.txt).
Environment fixtures derive from
[AgentDojo](https://github.com/ethz-spylab/agentdojo), with the original MIT
copyright notice retained in [licenses/AGENTDOJO-MIT.txt](licenses/AGENTDOJO-MIT.txt).
Hybrid50 adds contextual requests, candidate records, contracts, and reference
plans to those fixtures; it is not an official AgentDojo task release.
Original Hybrid50 contributions and release utilities are available under the
MIT terms in [LICENSE](LICENSE). Upstream components retain their respective
licenses; third-party retrieved pages retain their source attribution and any
applicable rights. See [DATA_LICENSE.md](DATA_LICENSE.md).
## Local verification and publication
```bash
python scripts/verify_release.py
pip install -r requirements.txt
python scripts/verify_release.py --check-parquet
python scripts/upload.py anonymous-quantinv/hybrid50
```
The canonical task file has byte SHA-256
`5e5b57fc5b1bba98e59f50fe765cdfed881769e2622ae19d59802f2781187aec`.
The replay report instead hashes the parsed record array serialized with
`json.dumps(records, sort_keys=True, ensure_ascii=False)`, yielding
`efb39403755a50fd6643b9171f0331143744ef9c547d2c853826e1c32cf998a9`.
Both hash definitions are recorded explicitly in the manifest.