fix: Change max_completion_tokens to max_tokens in io.yaml conversion script.

_ollama/convert_io_yaml_files.py

@@ -8,7 +8,6 @@ import copy
 import json
 import yaml
 
-
 # No automated way, just add mappings here as needed.
 MAP_MODELS_HF_TO_OLLAMA = {
     "granite-3.3-8b-instruct": "granite3.3:8b",
@@ -18,12 +17,16 @@ MAP_COLON = "_"
 LORA_TYPES = ["lora", "alora"]
 IO_FILE_NAME = "io.yaml"
 
-
 def find_all_model_paths(model_name: str) -> List[Path]:
     """Find all paths with the given model name.
     """
     current = Path(".")
-    return sorted([p for p in current.rglob(model_name) if p.is_dir()])
+    paths = []
+    for p in current.rglob(model_name):
+        rp = p.relative_to(current).parts
+        if p.is_dir() and not (rp[0].startswith("_") or rp[0].startswith("_")):
+            paths.append(p)
+    return sorted(paths)
 
 def map_model_path_hf_to_ollama(model_name: str, hf_path: Path) -> Path:
     """Replace HF model name in the model path to Ollama model name.
@@ -52,6 +55,12 @@ def convert_io_yaml_hf_to_ollama(hf_path: Path, ollama_path: Path):
     # Set movement of documents into message roles
     ollama_yaml["docs_as_message"] = "roles"
 
+    # Change to max_tokens as max_completion_tokens is not yet supported:
+    # https://github.com/ollama/ollama/issues/7125
+    if "max_completion_tokens" in ollama_yaml["parameters"]:
+        ollama_yaml["parameters"]["max_tokens"] = ollama_yaml["parameters"]["max_completion_tokens"]
+        del ollama_yaml["parameters"]["max_completion_tokens"]
+
     with ollama_path.open("w") as f:
         yaml.dump(ollama_yaml, f, default_style=False)
 
@@ -79,7 +88,6 @@ def convert_lora_adapters(model_name: str, model_paths: List[Path]):
                 ollama_lora_path.mkdir(parents=True, exist_ok=True)
                 convert_io_yaml_hf_to_ollama(hf_io_file_path, ollama_io_file_path)
 
-
 def main():
     """Main function.
     """
@@ -97,6 +105,7 @@ def main():
         )
 
     model_paths = find_all_model_paths(model_name)
+    print(model_paths)
 
     print(f"Found {len(model_paths)} intrinsics for {model_name}:")
     for model_path in model_paths:

answer_relevance_classifier/granite4_micro/lora/io.yaml

@@ -2,7 +2,7 @@ docs_as_message: roles
 instruction: answer_relevance
 model: null
 parameters:
-  max_completion_tokens: 1024
+  max_tokens: 1024
   response_format:
     properties:
       answer_relevance_analysis:

answer_relevance_rewriter/granite4_micro/lora/io.yaml

@@ -14,7 +14,7 @@ instruction: "Rewrite the response for relevance.\nThe last assistant response i
   \ their inquiry, in place of the original response.\n"
 model: null
 parameters:
-  max_completion_tokens: 1024
+  max_tokens: 1024
   response_format:
     properties:
       answer_relevance_rewrite:

answerability/granite4_micro/alora/io.yaml

@@ -2,7 +2,7 @@ docs_as_message: roles
 instruction: null
 model: null
 parameters:
-  max_completion_tokens: 6
+  max_tokens: 6
   response_format:
     enum:
     - answerable

answerability/granite4_micro/lora/io.yaml

@@ -2,7 +2,7 @@ docs_as_message: roles
 instruction: null
 model: null
 parameters:
-  max_completion_tokens: 6
+  max_tokens: 6
   response_format:
     enum:
     - answerable

citations/granite4_micro/lora/io.yaml

@@ -8,7 +8,7 @@ instruction: 'Split the last assistant response into individual sentences.  For
   '
 model: null
 parameters:
-  max_completion_tokens: 4096
+  max_tokens: 4096
   response_format:
     $defs:
       _MODEL_OUTPUT_ENTRY:

hallucination_detection/granite4_micro/lora/io.yaml

@@ -9,7 +9,7 @@ instruction: 'Split the last assistant response into individual sentences. For e
   '
 model: null
 parameters:
-  max_completion_tokens: 4096
+  max_tokens: 4096
   response_format:
     $defs:
       HallucinationOutputEntry:

query_rewrite/granite4_micro/lora/io.yaml

@@ -2,7 +2,7 @@ docs_as_message: roles
 instruction: null
 model: null
 parameters:
-  max_completion_tokens: 1024
+  max_tokens: 1024
   response_format:
     properties:
       rewritten_question: