Hugging Face's logo

ibm-granite
/
granitelib-rag-gpt-oss-r1.0

Safetensors
Model card Files
xet
 Community
12

Updates for use in granite-common - strip channel tokens

#2
by kndtran - opened
base: refs/heads/main
←
from: refs/pr/2
Discussion Files changed
+12
-3
kndtran
No description provided.
add strip_channel_tokens and increase answerability max tokens9ef495e4
kndtran

Summary

Add strip_channel_tokens: true to all GPT-OSS intrinsics and increase answerability max_completion_tokens from 6 to 25.

Changes

  • strip_channel_tokens: GPT-OSS models wrap their output in channel tokens (<|channel|>, <|message|>, <|return|>) which break JSON parsing. Added strip_channel_tokens: true to all four intrinsics:

    • answerability
    • citations
    • hallucination_detection
    • query_rewrite

  • max_completion_tokens: Increased from 6 to 25 for answerability to account for the additional channel token overhead.

Requires granite-common PR: https://github.com/ibm-granite/granite-common/pull/123

kndtran changed pull request status to open
kndtran changed pull request title from Updates for use in granite-common to Updates for use in granite-common - strip channel tokens
kndtran changed pull request status to closed
kndtran

Close in favor of https://huggingface.co/ibm-granite/granite-lib-rag-gpt-oss-r1.0/discussions/3

Sign up or log in to comment