framework
active
framework:data-centric-post-training-auditing-pipelineData-centric post-training auditing pipeline
Pipeline from preference data to hypotheses to interventions.
Neighborhood — ranked by edge-count
Papers (1)
paper
Methods (2)
method
- Finds data clusters where response concept is chosen/rejected leaning.
- Finds response-delta clusters shifted for prompt clusters.
Artifacts (1)
artifact
- Interface for inspecting prompt- and feature-conditioned hypotheses.
Quotes (1)
quote
- Core goal statement.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Iterative approach to construct challenging synthetic multi-hop QA pairs, long-form report writing tasks, and math/code reasoning tasks that exceed difficulty of existing datasets.
- Approach emphasizing data quality and source identification rather than only model architecture changes.
- Alignment approach that focuses on curating or modifying training data; the paper bridges this with interpretability methods.
- Broader research area: methods to align model behavior after initial training, where undesired behaviors can emerge.
- The phase after pre-training where models are further tuned with techniques like DPO; the period where the studied behavior emerged.
- Closing thesis.
- Downstream task validating NLA utility for model auditing; agents succeed without access to misalignment training data.