asingh15/glm52-aligned-rubric-traces

GLM-5.2 Aligned Rubric-Writing Traces

23777 teacher traces from GLM-5.2 on the aligned rubric-writing task, collected to distill / warmstart a smaller rubric-writer. For each (user, book) example the teacher is shown a persona-conditioned prompt (a user's past book reviews) and asked to (1) predict what that user would likely write about a new book and (2) produce a <rubric> of numbered criteria for scoring candidate reviews on coverage of that prediction. The full generation — thinking + rubric — is saved.

Fields

• —uid: reviewer id (groups traces by user)
• —index, sample: source row index and sample number (N=1: one generation per prompt)
• —writer_messages: the chat prompt sent to the teacher ([{"role":"user","content": ...}])
• —reasoning: the teacher's thinking (chain-of-thought)
• —rubric_raw: the teacher's final content (the <rubric>...</rubric> block)
• —rubric: the extracted rubric body
• —finish_reason: decoding stop reason (all stop)
• —teacher: teacher model tag (glm52b = GLM-5.2-FP8)

Generation

Training-matched decoding: temperature=1.0, top_p=1.0, thinking enabled, up to 32768 tokens.

Notes

The source-dataset gold (ground_truth: target reviews + candidate pools) is excluded. Traces are model outputs and may contain errors; intended for research (distillation/SFT).