choices null | id stringclasses 1
value | input stringclasses 1
value | metadata dict | target stringclasses 1
value |
|---|---|---|---|---|
null | concepts/agent-registries | Which ERC defines onchain identity, reputation, and validation registries for AI agents? Reply with just the number. | {
"eval_hash": "25f3d1726d05d2833f5aa4a617eb939112b67ce1b26d5729bdf09de18d371886",
"eval_id": "concepts/agent-registries",
"pillar": "concepts"
} | 8004 |
ETH Evals
This dataset contains 1 Ethereum quiz evals that declare the vanilla mode.
Each row holds one prompt and its target. Builds and agent modes stay in the source repository.
Dataset license: mit.
id identifies the eval. input, target, and choices use Inspect's sample fields.
metadata holds eval_id, pillar, and eval_hash. The hash covers every file in the source eval folder.
The export contains no logs, workspace files, or build solutions.
Each config groups one pillar and one set of scorer settings.
Names spell out the settings. Only pattern configs include a regex hash. The first config is the default.
eval.yaml names the solver and scorer for each config.
Match tasks set their location explicitly. Choice tasks use multiple_choice and choice.
The export rejects alternative target lists because Inspect's HF loader changes their meaning.
A single-item target list becomes its string value.
Load a config without a field mapping:
from inspect_ai.dataset import hf_dataset
dataset = hf_dataset("technophile-04/ethevals-demo", name="concepts-match-end", split="test")
Install Inspect's optional datasets dependency to use hf_dataset.
For a local copy, json_dataset("data/concepts-match-end/test.jsonl") needs no extra package.
Choose a config from the card's configs list. Each config has one test split.
Pass revision to hf_dataset to pin a published version.
After publication, run all benchmark tasks with stock Inspect:
inspect eval hf/technophile-04/ethevals-demo --model YOUR_PROVIDER/YOUR_MODEL
That command calls the selected model. It needs the provider's credentials and can incur charges.
Install huggingface_hub and datasets for the hf/ task loader.
The hf/ loader reads the same prompt, choices, target, and scoring settings.
Inspect 0.3.271 assigns sample IDs on that path and nests the source metadata under metadata.
Direct hf_dataset and json_dataset loads retain the row ID and metadata without a mapping.
These public targets can appear in training data. Scores on them alone do not establish unseen Ethereum knowledge.
- Downloads last month
- 41