code/slurm_logs/mega_expand_datasetC.sbatch

#!/bin/bash
#SBATCH --job-name=vl
#SBATCH --partition=gpu-large
#SBATCH --qos=batch-long
#SBATCH --gres=gpu:1
#SBATCH --cpus-per-task=4
#SBATCH --mem=80G
#SBATCH --time=08:00:00
#SBATCH --output=/weka/s225250685/mats-tist/slurm_logs/expand_datasetC_%j.out

# Expand Dataset C: serve thanhdath planner, run greedy K=1 on uncovered BIRD train
# questions, extract correct CoT, pair with griffith prompts → append to Dataset C.
# Currently 1106 covered → target 4000+ questions (46% greedy accuracy × 8322 uncovered)

set -u
cd /weka/s225250685/mats-tist

export HF_HOME=/weka/s225250685/Huggingface
export HF_HUB_CACHE=/weka/s225250685/Huggingface/hub
export DB_EXEC_API_DISABLE=1
export PYTHONNOUSERSITE=1
export NO_PROXY=localhost,127.0.0.1
export PYTHONPATH=/weka/s225250685/mats-tist
export TOKENIZERS_PARALLELISM=false

PY=/weka/s225250685/conda-envs/handbook/bin/python
VLLM=/weka/s225250685/conda-envs/handbook/bin/vllm
THANHDATH=/weka/s225250685/Huggingface/hub/models--thanhdath--orpo-llama-3b-iter-2-bird-planner/snapshots/8171b8585a306709996796b86de19c3dd39a910c
LOG=/weka/s225250685/mats-tist/slurm_logs/expand_datasetC_${SLURM_JOB_ID}.log
: > "$LOG"

nvidia-smi --query-gpu=name,memory.total --format=csv,noheader | tee -a "$LOG"

kill_vllm() { pkill -9 -f "vllm serve" 2>/dev/null || true; sleep 3; }
trap kill_vllm EXIT

##############################################
# STAGE A: Serve thanhdath planner
##############################################
echo "==== [A] serving thanhdath planner ====" | tee -a "$LOG"
$VLLM serve "$THANHDATH" --served-model-name planner --port 8100 \
    --dtype bfloat16 --gpu-memory-utilization 0.85 \
    --enforce-eager --max-model-len 8192 > "${LOG}.p" 2>&1 &

for i in {1..180}; do
    curl --noproxy '*' -fs http://localhost:8100/v1/models >/dev/null 2>&1 && break; sleep 5
done
echo "  planner READY" | tee -a "$LOG"

##############################################
# STAGE B: Run greedy K=1 on uncovered BIRD train questions
#   Prompt format: old dict schema (what thanhdath was trained on)
#   Output: correct (prompt_b, completion_a) pairs for Dataset C
##############################################
echo "==== [B] expanding Dataset C ====" | tee -a "$LOG"

$PY - << 'PYEOF' 2>&1 | tee -a "$LOG"
import json, os, re, random, sqlite3, threading, requests
from datasets import load_dataset, load_from_disk, Dataset, DatasetDict

ROOT = "/weka/s225250685/mats-tist"
os.chdir(ROOT)
HF_CACHE = "/weka/s225250685/Huggingface/hub"

# Load existing Dataset C to know which questions are already covered
existing = load_from_disk("data/hf_planner_sft_griffith")
covered_questions = set()
for split in ["train","test"]:
    for row in existing[split]:
        covered_questions.add(row["question"].strip().lower())
print(f"Already covered: {len(covered_questions)} questions", flush=True)

# Load griffith prompts (all 9428)
with open("data/sft_bird_with_evidence_train_text2sql.json") as f:
    bird_train = json.load(f)

ds_b = load_dataset("griffith-bigdata/sft_text2sql", split="train_sft",
                    cache_dir=HF_CACHE).filter(lambda x: x["model_name"]=="deepseek-reasoner")

# Build griffith prompt lookup: question_lower → (prompt_b, sid, db_id)
griffith_lookup = {}
for row in ds_b:
    sid = int(row["sample_id"])
    if sid >= len(bird_train): continue
    user_msg = row["messages"][1]["content"]
    q_m = re.search(r"Question:\s*(.+?)(?:\n|$)", user_msg)
    if not q_m: continue
    griffith_q = q_m.group(1).strip()
    bird_q = bird_train[sid]["question"].strip()
    if griffith_q.lower() != bird_q.lower(): continue
    q_key = bird_q.lower()
    if q_key not in covered_questions:
        griffith_lookup[q_key] = {
            "prompt_b": user_msg.rstrip() + "\n\nPlanning:",
            "sid": sid, "db_id": bird_train[sid].get("db_id",""),
            "question": bird_q, "gold_sql": bird_train[sid]["sql"],
            "db_path": bird_train[sid].get("db_path",""),
        }

print(f"Uncovered questions with griffith prompts: {len(griffith_lookup)}", flush=True)

# Planner inference helpers
def llama3_chat(p):
    return (f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n"
            f"{p}<|eot_id|>\n<|start_header_id|>assistant<|end_header_id|>\n")

def safe_exec(db_path, sql, timeout=5):
    r=[None]; e=[None]
    def _run():
        try:
            c=sqlite3.connect(db_path); c.text_factory=lambda b:b.decode(errors="ignore")
            r[0]=c.execute(sql).fetchmany(100); c.close()
        except Exception as ex: e[0]=str(ex)
    t=threading.Thread(target=_run,daemon=True); t.start(); t.join(timeout)
    return (None,"TIMEOUT") if t.is_alive() else (r[0],e[0])

def results_match(gold, pred):
    if gold is None or pred is None: return False
    def norm(rows):
        return sorted(tuple(str(v).strip().lower() if v is not None else "" for v in row) for row in rows)
    return norm(gold) == norm(pred)

def extract_sql(text):
    m = re.search(r"```(?:sql)?\s*(.*?)\s*```", text, re.DOTALL)
    if m:
        sql = m.group(1).strip()
        return sql[3:].strip() if sql.upper().startswith("SQL") else sql
    lines = [l.strip() for l in text.strip().split("\n") if l.strip()]
    return lines[-1] if lines else ""

PLANNER_PROMPT_TMPL = "{schema}\n\nQuestion: {question}\nExternal knowledge: {evidence}\n\nPlanning:"

new_rows = []
n_correct = 0; n_wrong = 0

items = list(griffith_lookup.values())
random.seed(42); random.shuffle(items)

for i, info in enumerate(items):
    sid = info["sid"]
    bt = bird_train[sid]
    # Build OLD schema prompt for thanhdath planner
    old_schema = str(bt.get("schema_sequence") or bt.get("schema") or "")
    old_prompt = PLANNER_PROMPT_TMPL.format(
        schema=old_schema, question=bt["question"],
        evidence=bt.get("evidence","") or "None",
    )
    raw_prompt = llama3_chat(old_prompt)
    try:
        r = requests.post("http://localhost:8100/v1/completions", json={
            "model": "planner", "prompt": raw_prompt,
            "max_tokens": 1024, "temperature": 0.0, "n": 1,
            "seed": 42, "stop": ["<|eot_id|>"],
        }, timeout=30)
        r.raise_for_status()
        cot = r.json()["choices"][0]["text"].strip()
    except Exception as ex:
        n_wrong += 1; continue

    pred_sql = extract_sql(cot)
    if not pred_sql:
        n_wrong += 1; continue

    db_path = info["db_path"] or f"data/train_databases/{info['db_id']}/{info['db_id']}.sqlite"
    gold_res, _ = safe_exec(db_path, info["gold_sql"])
    pred_res, err = safe_exec(db_path, pred_sql)

    if err or not results_match(gold_res, pred_res):
        n_wrong += 1; continue

    # Correct → pair griffith prompt_b with this CoT completion_a
    new_rows.append({
        "prompt":     info["prompt_b"],    # griffith NL schema + Planning:
        "completion": cot,                  # CoT from thanhdath (correct)
        "sample_id":  sid,
        "db_id":      info["db_id"],
        "question":   info["question"],
    })
    n_correct += 1

    if (i+1) % 500 == 0:
        acc = n_correct/(n_correct+n_wrong)*100 if (n_correct+n_wrong) else 0
        print(f"  [{i+1}/{len(items)}] correct={n_correct} acc={acc:.1f}%", flush=True)

print(f"\nNew pairs: {len(new_rows)} correct / {n_correct+n_wrong} attempted", flush=True)

# Merge with existing Dataset C and resave
all_rows = []
for split in ["train","test"]:
    for row in existing[split]:
        all_rows.append({k: row[k] for k in ["prompt","completion","sample_id","db_id","question"]})
all_rows.extend(new_rows)

random.shuffle(all_rows)
n_train = int(0.9 * len(all_rows))
DatasetDict({
    "train": Dataset.from_list(all_rows[:n_train]),
    "test":  Dataset.from_list(all_rows[n_train:]),
}).save_to_disk("data/hf_planner_sft_griffith_expanded")

print(f"Saved → data/hf_planner_sft_griffith_expanded  (train={n_train}, test={len(all_rows)-n_train})", flush=True)
PYEOF

kill_vllm
echo "==== ALL_DONE ====" | tee -a "$LOG"