Upload script.py with huggingface_hub

script.py

@@ -1,7 +1,27 @@
+import subprocess
+import sys
+
+# Eval sandbox ships older transformers/tokenizers; tiny-aya needs recent versions.
+# Pip has network even though model loading is offline.
+subprocess.run(
+    [
+        sys.executable,
+        "-m",
+        "pip",
+        "install",
+        "-q",
+        "transformers>=4.56",
+        "tokenizers>=0.21.1",
+        "accelerate>=0.30",
+        "torch>=2.2",
+    ],
+    check=True,
+)
+
 import os
 import re
+import csv
 import json
-import pandas as pd
 import torch
 from transformers import AutoTokenizer, AutoModelForCausalLM
 
@@ -46,10 +66,11 @@ model = AutoModelForCausalLM.from_pretrained(
     MODEL_ID, torch_dtype=torch.float16, device_map="auto"
 ).eval()
 
-df = pd.read_csv("/tmp/data/test.csv", dtype=str).fillna("")
+with open("/tmp/data/test.csv", encoding="utf-8", newline="") as f:
+    test_rows = list(csv.DictReader(f))
 
 outputs_queries_types = []
-for _, r in df.iterrows():
+for r in test_rows:
 
     # Create the prompt.
     messages = [
@@ -81,7 +102,7 @@ for _, r in df.iterrows():
             done = True
 
     outputs_queries_types.append((text, r['id'], r['query'], r['task_type']))
-    print(f"{len(outputs_queries_types)}/{len(df)} done", flush=True)
+    print(f"{len(outputs_queries_types)}/{len(test_rows)} done", flush=True)
 
 # Postprocess and store the answers.
 def expected_answer_count(query: str, task_type: str) -> int:
@@ -200,5 +221,8 @@ rows = []
 for answer, row_id, query, task_type in outputs_queries_types:
     answers = postprocess_answer(answer, query, task_type)
     rows.append({"id": row_id, "pred": json.dumps(answers, ensure_ascii=False)})
-pd.DataFrame(rows).to_csv("submission.csv", index=False)
+with open("submission.csv", "w", encoding="utf-8", newline="") as f:
+    writer = csv.DictWriter(f, fieldnames=["id", "pred"])
+    writer.writeheader()
+    writer.writerows(rows)
 print("wrote submission.csv", flush=True)