verifier-rm-data / README.md
juppy44's picture
Format rows as language-model conversations
3712700 verified
|
Raw
History Blame Contribute Delete
760 Bytes

Verifier RM data — conversation format

This revision contains the verifier reward-model examples formatted for language-model conversation training.

Each row has one column named messages with two turns:

  • user: instruction, current completion, candidate text, and optional evidence
  • assistant: the original target_label (YES or NO)

The user content uses this order: <instruction>, <current_completion>, and <text_to_score>. Non-empty reference_solution and privileged_context values are appended independently. The requested privelaged_context spelling is retained in the tag.

Split counts are train 48,607, validation 2,697, and test 2,651.

Source revision used for this conversion: 556b271958cd5c6afb2093b5333b64dba5c42d6a.