team-riswan / training /kaggle_cells /cell1_pass1.py
riswanmsm's picture
Submit final system code, manifest, and predictions for team-riswan
682bee1
Raw History Blame Contribute Delete
6.3 kB
import json, re, time, zipfile, shutil, urllib.request
from pathlib import Path
from concurrent.futures import ThreadPoolExecutor, as_completed
# 1. Locate uploaded data in /kaggle/input and copy to /kaggle/working/work
work_dir = Path("/kaggle/working/work")
work_dir.mkdir(parents=True, exist_ok=True)
for p in Path("/kaggle/input").rglob("*"):
if p.name == "kaggle_data.zip":
with zipfile.ZipFile(p, "r") as z:
z.extractall(work_dir)
break
elif p.name == "predictions_gemma4_cot.jsonl":
root = p.parent.parent
shutil.copytree(root, work_dir, dirs_exist_ok=True)
break
data_dir = work_dir / "data"
print("Data directory ready:", data_dir.exists())
# 2. Load canonical questions + existing checkpoints
def load_all_questions(d_dir: Path):
qs = []
for lvl in ["easy", "medium", "hard"]:
for f in sorted((d_dir / lvl).glob("*.json")):
for item in json.load(open(f, encoding="utf-8")):
item["_level"] = lvl
qs.append(item)
return qs
def load_ckpt(path: Path):
m = {}
if path.exists():
for line in open(path, encoding="utf-8"):
if line.strip():
r = json.loads(line)
k = f"{str(r['level']).lower().strip()}_{str(r['subject']).strip()}_{str(r['q_no']).strip()}"
m[k] = r
return m
canonical_qs = load_all_questions(data_dir)
cot_map = load_ckpt(data_dir / "predictions_gemma4_cot.jsonl")
dir_map = load_ckpt(data_dir / "predictions_gemma4_direct.jsonl")
# Sort: 749 Disputed (Hard -> Medium -> Easy) first, then Undisputed (Hard -> Medium -> Easy)
lvl_rank = {"hard": 0, "medium": 1, "easy": 2}
def sort_key(q):
k = f"{q['_level'].lower()}_{q['subject'].strip()}_{q['q_no']}"
disputed = 0 if cot_map.get(k, {}).get("answer") != dir_map.get(k, {}).get("answer") else 1
return (disputed, lvl_rank.get(q["_level"].lower(), 3))
priority_qs = sorted(canonical_qs, key=sort_key)
print(f"Loaded {len(canonical_qs)} total questions. Starting Dual-T4 Parallel Thinking Run...")
SYSTEM_THINK = (
"You are an expert teacher of the Sri Lankan national school curriculum (grades 6-13), "
"fluent in Sinhala. You will receive a multiple-choice question written in Sinhala with four "
"numbered options. Exactly one option is correct. Read the Sinhala carefully, including "
"negations such as 'නොවන' / 'නොවේ' (NOT). You may reason in English. "
"Finish with the final answer in the form 'Answer: <number>'."
)
ANS_RE = re.compile(r"(?:Answer|පිළිතුර)\s*[::]?\s*\**\s*\(?([1-4])\)?", re.I)
HOSTS = ["http://127.0.0.1:11434", "http://127.0.0.1:11435"]
def solve_question(idx_q):
idx, q = idx_q
host = HOSTS[idx % 2] # Alternate evenly between T4 GPU #0 and T4 GPU #1
lvl = q["_level"].lower()
subj = q.get("subject", "General")
q_no = q["q_no"]
q_key = f"{lvl}_{subj.strip()}_{q_no}"
grade = "6-9" if lvl == "easy" else ("10-11 (O/L)" if lvl == "medium" else "12-13 (A/L)")
opts = "\n".join(f"{i}. {c}" for i, c in enumerate(q["choices"], 1))
user_msg = (
f"විෂය (Subject): {subj} | ශ්‍රේණිය (Grade): {grade}\n\n"
f"ප්‍රශ්නය (Question):\n{q['question'].strip()}\n\n"
f"විකල්ප (Options):\n{opts}\n\n"
"Which option (1, 2, 3 or 4) is correct? End your reply with 'Answer: <number>'."
)
payload = json.dumps({
"model": "gemma4:e4b",
"messages": [{"role": "system", "content": SYSTEM_THINK}, {"role": "user", "content": user_msg}],
"stream": False,
"think": True,
"options": {"temperature": 0.0, "num_predict": 2048}
}).encode("utf-8")
try:
req = urllib.request.Request(f"{host}/api/chat", data=payload, headers={"Content-Type": "application/json"})
with urllib.request.urlopen(req, timeout=240) as resp:
res = json.loads(resp.read().decode("utf-8"))
msg = res.get("message", {})
content, thinking = msg.get("content", ""), msg.get("thinking", "")
m = ANS_RE.findall(content) or ANS_RE.findall(thinking)
if m:
ans = m[-1]
else:
digits = re.findall(r"\b([1-4])\b", content if content.strip() else thinking)
ans = digits[-1] if digits else cot_map.get(q_key, {}).get("answer", "1")
except Exception as e:
ans = cot_map.get(q_key, {}).get("answer", "1")
return q_key, {
"q_no": q_no,
"subject": subj,
"category": q.get("category", "general"),
"level": lvl,
"answer": str(ans)
}
ckpt_path = Path("/kaggle/working/predictions_gemma4_think_ckpt.jsonl")
full_path = Path("/kaggle/working/predictions.jsonl")
zip_path = Path("/kaggle/working/submission_gemma4_think.zip")
completed_map = load_ckpt(ckpt_path)
todo = [(i, q) for i, q in enumerate(priority_qs) if f"{q['_level'].lower()}_{q['subject'].strip()}_{q['q_no']}" not in completed_map]
print(f"Already completed: {len(completed_map)} | Remaining to solve: {len(todo)}")
def save_outputs():
with open(ckpt_path, "w", encoding="utf-8") as f:
for r in completed_map.values():
f.write(json.dumps(r, ensure_ascii=False) + "\n")
with open(full_path, "w", encoding="utf-8") as f:
for q in canonical_qs:
k = f"{q['_level'].lower()}_{q['subject'].strip()}_{q['q_no']}"
rec = completed_map.get(k) or cot_map.get(k)
f.write(json.dumps(rec, ensure_ascii=False) + "\n")
with zipfile.ZipFile(zip_path, "w", zipfile.ZIP_DEFLATED) as zf:
zf.write(full_path, arcname="predictions.jsonl")
t0 = time.time()
done_count = 0
with ThreadPoolExecutor(max_workers=4) as ex:
futures = [ex.submit(solve_question, item) for item in todo]
for fut in as_completed(futures):
k, rec = fut.result()
completed_map[k] = rec
done_count += 1
if done_count % 20 == 0 or done_count == len(todo):
save_outputs()
elapsed = time.time() - t0
rate = elapsed / done_count
print(f"[{len(completed_map)}/2238] ({elapsed/60:.1f}m elapsed, {rate:.1f}s/q effective) -> {zip_path.name} updated!")