fix: pre-render chat template via .map() instead of formatting_func

train_lora.py

@@ -58,30 +58,29 @@ def main() -> None:
         random_state=42,
     )
 
-    # Load dataset and split off a tiny eval slice for live monitoring during training
+    # Load dataset, pre-render the chat template into a single "text" column,
+    # then split. Avoids version-skew on TRL/Unsloth's formatting_func contracts —
+    # SFTTrainer reads dataset_text_field="text" and tokenizes directly.
     full = load_dataset(DATASET_REPO, data_files="train.jsonl", split="train")
-    split = full.train_test_split(test_size=0.05, seed=42)
 
-    # Unsloth's wrapper invokes formatting_func with a BATCH (dict of lists) and
-    # expects a list-of-strings back — NOT TRL's per-example shape. Iterate the
-    # batch and apply the chat template to each turn-list.
-    def formatting_func(examples: dict) -> list[str]:
-        return [
-            tokenizer.apply_chat_template(
-                msgs,
+    def render_chat(example: dict) -> dict:
+        return {
+            "text": tokenizer.apply_chat_template(
+                example["messages"],
                 tokenize=False,
                 add_generation_prompt=False,
             )
-            for msgs in examples["messages"]
-        ]
+        }
+    full = full.map(render_chat, remove_columns=["messages"])
+    split = full.train_test_split(test_size=0.05, seed=42)
 
     trainer = SFTTrainer(
         model=model,
         tokenizer=tokenizer,
         train_dataset=split["train"],
         eval_dataset=split["test"],
-        formatting_func=formatting_func,
         args=SFTConfig(
+            dataset_text_field="text",
             output_dir="raunch-stheno-v3.4-lora-v0",
             push_to_hub=True,
             hub_model_id=MODEL_REPO,