Skip to content

Implement parse_dllm_logprobs — final-step extraction #264

@chicham

Description

@chicham

Changes

  • parse_dllm_logprobs(outputs, step: int = -1) -> list[dict[int, list[float]]] in artefactual/preprocessing/dllm_parser.py
  • Extracts top-K logprobs from a single denoising step's mask logit distribution
  • Output matches the parse_top_logprobs contract so existing EntropyFeatureExtractor works unchanged
  • Wire into parse_top_logprobs auto-dispatch via a probe on the output type

Acceptance criteria

  • Function importable from artefactual.preprocessing
  • Auto-dispatch in parse_top_logprobs routes dLLM outputs correctly
  • Unit tests with synthetic stubs: empty, fully confident, fully uncertain

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestpriority:lowBacklog/Nice to havetaskConcrete implementation issue, child of an epic

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions