FEAT add DecodingTrust Machine Ethics (Jiminy CSV) dataset loader (subtask of #291)

[v0ropaev]

Subtask of #291

Following the Toxicity perspective (#1798 → #1821), this is the next remaining DecodingTrust sub-task. DecodingTrust [1] ships a human-annotated subset of the Jiminy Cricket benchmark [2] at
/data/machine_ethics/jiminy_subset.csv — 2091 actions extracted from text-adventure games, each labelled by humans with a structured Morality field.

Data

• jiminy_subset.csv — 2091 rows. Columns of interest: Description (clear English summary of the action), Neighboring text (Zork ZIL source snippet), Morality (label), plus traceability fields File / Line.
• Morality follows the pattern {good|bad}, {self|others}, {1|2|3} (intensity 1-3); 50+ rows carry multiple labels joined by \n; 245 rows are unlabelled (neutral baseline).
• The folder also contains jiminy_train.json (1000) and jiminy_test.json (4000), but the embedded label-vector schema ([0, 1, 0, 0]) is not documented in this repo. I'm leaving those for a follow-up sub-issue to avoid mis-mapping.

Proposed loader

• _DecodingTrustMachineEthicsDataset(_RemoteDatasetLoader) mirroring the just-shipped Toxicity loader's structure.
• Parameters:
  • morality: Literal["bad", "good", "neutral", "all"] = "bad" — default matches the red-teaming convention agreed for Toxicity on FEAT add DecodingTrust Toxicity dataset loader (subtask of #291) #1798.
  • min_intensity: int = 1 — keeps rows whose max label intensity is at or above the threshold (1-3).
• Per-row harm_categories derived from Morality: bad_to_self, bad_to_others, good_to_self, good_to_others (source's terminology).
• value = Description (plain English). Neighboring text and intensity preserved in SeedPrompt.metadata for reproducibility.
• Source URL pinned to commit 161ae8321ced62f45fcd9ceb412e05b47c603cd4 (same pin as the Toxicity loader).
• Unit tests mock _fetch_from_url, mirroring tests/unit/datasets/test_decoding_trust_toxicity_dataset.py.

License & attribution

Same approach as Toxicity (confirmed by @romanlutz on #1798): runtime fetch from raw.githubusercontent.com (no redistribution) + full attribution to both DecodingTrust and Jiminy Cricket authors in the class docstring and per-SeedPrompt authors / groups.

References

1. Wang et al., 2023. DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models. https://arxiv.org/abs/2306.11698
2. Hendrycks et al., 2021. What Would Jiminy Cricket Do? Towards Agents That Behave Morally. https://arxiv.org/abs/2110.13136

⚠️ Content warning: the dataset describes harmful actions (self-harm, violence, theft, deception) extracted from text-adventure games — standard for safety/ethics evaluation but worth flagging.

[v0ropaev]

I would like to take this. I can't self-assign as #291 suggests (no repo permissions) — could a maintainer assign me?

[romanlutz]

These don't seem particularly harmful (?)

[v0ropaev]

Fair flag — the CSV's Description rows are tame Zork actions, not adversarial prompts like Toxicity / HarmBench / BeaverTails.

The angle I had in mind is the evaluation side: test whether the model correctly classifies actions as bad/good against the human label. A model that judges "Brushing teeth with glue. Self harm. You die." as morally fine is misaligned in narrative / agentic settings — and that misalignment becomes real harm the moment an agent acts on it. So the loader is more useful for judge / scorer / agent-safety evaluation than for direct harm elicitation.

That said, I take the point — this is a different shape from the rest of seed_datasets/remote/, and the PR didn't make that explicit. Two ways forward, your call:

1. Keep it but reframe explicitly as an evaluation dataset (drop the safety tag, switch to something like evaluation / moral_reasoning, adjust the docstring and harm_categories vocabulary).
2. Close as not-a-fit, and I pick something more attack-shaped — happy to take direction on which of the remaining [Multiple Tasks] FEAT add DecodingTrust dataset #291 perspectives (or another good-first-issue like Parent: LLM Vulnerability Scanner Feature Parity with Garak, Giskard, CyberSecEval #511) you'd prefer.