Skip to content

[Triage] M-A18-049 as worded is wrong: 'truthfulqa' resolves in lm-eval 0.4.x, as a TAG (neither a task nor a group) (CR-A05-075) #823

Closed as not planned

Description

@cemililik

Needs triage. Reviewers disagreed on this finding, so it is not counted in the backlog. Decide whether it is a real defect: if yes, set severity and wave; if not, close it with the reason.

CR-A05-075 — M-A18-049 as worded is wrong: 'truthfulqa' resolves in lm-eval 0.4.x, as a TAG (neither a task nor a group)

severity Info · confidence disputed · single reviewer · reproduced end to end

What happens. Real lm_eval.tasks.TaskManager from the PyPI wheels 0.4.4, 0.4.9.2 and 0.4.13 indexes 'truthfulqa' with type 'tag' (0.4.13: Entry kind=Kind.TAG, tags={truthfulqa_mc1, truthfulqa_mc2, truthfulqa_gen}). No truthfulqa.yaml exists in lm_eval/tasks/truthfulqa/ (only truthfulqa_gen.yaml, truthfulqa_mc1.yaml, truthfulqa_mc2.yaml, utils.py, README.md). Each leaf declares 'tag:\n - truthfulqa' on lines 1-2 (truthfulqa_gen.yaml, truthfulqa_mc1.yaml); truthfulqa_mc2.yaml gets the tag through 'include: truthfulqa_mc1.yaml'. 'truthfulqa' is in TaskManager.all_tags and not in all_groups. lm-eval's own benchmarks/openllm.yaml uses '- task: truthfulqa'. So the name the site, the CLI help and the wizard hint use is accepted by simple_evaluate and cache-tasks. It is not a task, and it is not a group with aggregation. The claim 'does not exist as a task name' is true only in the narrow sense that it is a tag. The real defect is how ForgeLM scores the expansion (next findings). The 0.4.4 floor already uses 'tag:'; older 0.4.0-0.4.3 used 'group:' but are below the declared floor.

Where.

Evidence.

  • 0.4.4: {'type':'tag','task':['truthfulqa_mc1','truthfulqa_gen','truthfulqa_mc2']}, is_tag True, is_group False, is_task False. 0.4.9.2: same members. 0.4.13: Entry(kind=Kind.TAG, tags={...3}). Wheel sha256: 0.4.4 37f1ab84..4fc3, 0.4.9.2 b8d12410..6a6c, 0.4.13 5daaa197..286a. The YAML ('tag:' not 'group:') is in tasks/truthfulqa/*.yaml in each wheel.

Recommended fix. Restate the item: the name is valid. Report the scoring and naming consequences below instead.


Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.

Activity

  1. added
    bugSomething isn't working
    needs-triageNot triaged yet: every new issue starts with it; only a maintainer removes it
    source: review-2026-10Found by the October 2026 code and documentation review
    on Oct 9, 2026
  2. cemililik commented on Oct 20, 2026

    @cemililik
    CollaboratorAuthor

    Thank you for the thorough check. I read the lm-evaluation-harness source at v0.4.4 and v0.4.13 and it agrees: no truthfulqa.yaml exists, the three leaf tasks declare the tag 'truthfulqa', and the task manager expands tag names, so the name ForgeLM uses is valid.

    Nothing needs to change for the name itself, so this is closed as not a defect. The consequences you list (three equally weighted entries, truthfulqa_gen scored through the bleu_acc fallback, no 'truthfulqa' key in the results, docs showing one) stay open in #120 and #613, where the examples can move to explicit task names such as truthfulqa_mc2.

    Triage round 1, checked against development on 2026-10-20.

  3. removed
    needs-triageNot triaged yet: every new issue starts with it; only a maintainer removes it
    bugSomething isn't working
    on Oct 20, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    source: review-2026-10Found by the October 2026 code and documentation review

    Type

    No type

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions