Upload finetune_soc.py with huggingface_hub

finetune_soc.py

@@ -32,6 +32,14 @@ OUTPUT_REPO = "Colby/apertus-8b-soc"
 print("Loading tokenizer...")
 tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
 
+# Apertus-8B tokenizer has no chat_template set; define ChatML explicitly.
+# Custom roles (persona usernames) are preserved as-is in the im_start tag.
+tokenizer.chat_template = (
+    "{% for message in messages %}"
+    "<|im_start|>{{ message['role'] }}\n{{ message['content'] }}<|im_end|>\n"
+    "{% endfor %}"
+)
+
 print("Loading dataset...")
 dataset = load_dataset(DATASET_ID, split="train")
 print(f"Loaded {len(dataset)} conversations")