Add logprobs workaround for harmony channel tokens

answerability/gpt-oss-20b/lora/io.yaml

@@ -20,8 +20,11 @@ transformations:
 instruction: ~
 parameters:
   # "unanswerable" can be 6 tokens at high temperatures
-  max_completion_tokens: 6
+  # channel tokens add overhead on gpt-oss models
+  max_completion_tokens: 25
 # No sentence boundary detection
 sentence_boundaries: ~
 # RAG documents go in first message
 docs_as_message: string
+# Use logprob tokens as ground truth for model output to handle harmony channel tokens
+logprobs_workaround: true

citations/gpt-oss-20b/lora/io.yaml

@@ -95,3 +95,5 @@ sentence_boundaries:
   documents: "c"
 # gpt-oss base models have no "documents" argument
 docs_as_message: json
+# Use logprob tokens as ground truth for model output to handle harmony channel tokens
+logprobs_workaround: true

hallucination_detection/gpt-oss-20b/lora/io.yaml

@@ -78,4 +78,6 @@ sentence_boundaries:
   last_message: "i"
 
 # gpt-oss base model has no "documents" argument
-docs_as_message: json
+docs_as_message: json
+# Use logprob tokens as ground truth for model output to handle harmony channel tokens
+logprobs_workaround: true

query_rewrite/gpt-oss-20b/lora/io.yaml

@@ -20,3 +20,5 @@ instruction: ~
 parameters:
   max_completion_tokens: 1024
 sentence_boundaries: false
+# Use logprob tokens as ground truth for model output to handle harmony channel tokens
+logprobs_workaround: true