fix: manual JSONL loader — pyarrow chokes on mixed tool_calls types

train_v4.py

@@ -69,7 +69,6 @@ model = FastLanguageModel.get_peft_model(
 )
 
 # ── Dataset ─────────────────────────────────────────────────────────
-ds = load_dataset("json", data_files=DATA, split="train")
 
 
 def fix_tool_calls(messages):
@@ -95,35 +94,41 @@ def fix_tool_calls(messages):
     return fixed
 
 
-def to_chatml(ex):
+def load_and_format(path):
-    """Apply Qwen3.5 base chat template with thinking enabled."""
+    """Load JSONL manually — pyarrow chokes on mixed tool_calls argument types."""
-    messages = fix_tool_calls(ex["messages"])
+    from datasets import Dataset
-    try:
+    texts = []
-        text = tokenizer.apply_chat_template(
+    skipped = 0
-            messages,
+    with open(path) as f:
-            tokenize=False,
+        for line in f:
-            add_generation_prompt=False,
+            line = line.strip()
-            # Enable thinking mode so <think> blocks are properly formatted
+            if not line:
-            enable_thinking=True,
+                continue
-        )
+            row = json.loads(line)
-    except TypeError:
+            messages = fix_tool_calls(row["messages"])
-        # Fallback if enable_thinking not supported in this template version
+            try:
-        text = tokenizer.apply_chat_template(
+                text = tokenizer.apply_chat_template(
-            messages,
+                    messages,
-            tokenize=False,
+                    tokenize=False,
-            add_generation_prompt=False,
+                    add_generation_prompt=False,
-        )
+                    enable_thinking=True,
-    return {"text": text}
+                )
+            except TypeError:
+                text = tokenizer.apply_chat_template(
+                    messages,
+                    tokenize=False,
+                    add_generation_prompt=False,
+                )
+            if len(tokenizer.encode(text)) <= MAX_SEQ:
+                texts.append(text)
+            else:
+                skipped += 1
+    if skipped:
+        print(f"⚠ Filtered {skipped} examples exceeding {MAX_SEQ} tokens")
+    return Dataset.from_dict({"text": texts})
 
 
-ds = ds.map(to_chatml)
+ds = load_and_format(DATA)
-
-# Filter out examples that exceed max sequence length
-orig_len = len(ds)
-ds = ds.filter(lambda ex: len(tokenizer.encode(ex["text"])) <= MAX_SEQ)
-filtered = orig_len - len(ds)
-if filtered > 0:
-    print(f"⚠ Filtered {filtered} examples exceeding {MAX_SEQ} tokens")
 
 steps = (len(ds) * EPOCHS) // (BATCH * GRAD_ACCUM)
 print(f"Dataset: {len(ds)} examples")