code/slurm_logs/mega_planner_sft_v1.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/planner_sft_v1_%j.out

# ============================================================
# Planner SFT v1 — Dataset C: prompt_b (griffith rich NL) + completion_a (correct CoT)
# Base model: Qwen/Qwen2.5-Coder-3B-Instruct (NOT thanhdath ORPO model)
#
# Dataset C already built: data/hf_planner_sft_griffith
#   train=1877, test=209, 1106 unique questions
#   prompt format: griffith NL schema + "Planning:" trigger (Qwen chat format)
#   completion format: full CoT (Goal→Condition→Tables→Final SQL)
#
# Stage A: SFT fine-tune Qwen2.5-Coder-3B-Instruct on Dataset C
# Stage B: Quick oracle comparison vs baseline on 200 BIRD-dev questions
# ============================================================

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

BASE=Qwen/Qwen2.5-Coder-3B-Instruct
OUT_PLANNER=/weka/s225250685/mats-tist/alignment-handbook/output/planner-v1-qwen3b-griffith-sft

LOG=/weka/s225250685/mats-tist/slurm_logs/planner_sft_v1_${SLURM_JOB_ID}.log
: > "$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
    pkill -9 -f "VLLM::EngineCore" 2>/dev/null || true
    sleep 5
}
trap kill_vllm EXIT

wait_url() {
    for i in {1..180}; do
        curl --noproxy '*' -fs "$1" >/dev/null 2>&1 && return 0
        sleep 5
    done
    echo "TIMEOUT: $1" | tee -a "$LOG"; return 1
}

# Confirm Dataset C
$PY - << 'PYEOF' 2>&1 | tee -a "$LOG"
from datasets import load_from_disk
d = load_from_disk("data/hf_planner_sft_griffith")
print(f"Dataset C: train={len(d['train'])} test={len(d['test'])}")
ex = d['train'][0]
print(f"  db={ex['db_id']}  q={ex['question'][:60]}")
print(f"  prompt tail: ...{ex['prompt'][-80:]}")
print(f"  completion:  {ex['completion'][:100]}...")
PYEOF

##############################################
# STAGE A: SFT Qwen2.5-Coder-3B-Instruct on Dataset C
#   Qwen chat format: <|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n{CoT}<|im_end|>
##############################################
echo "==== [A] SFT Qwen2.5-Coder-3B on Dataset C ====" | tee -a "$LOG"

$PY - << 'PYEOF' 2>&1 | tee -a "$LOG"
import os, sys, torch
from transformers import (AutoModelForCausalLM, AutoTokenizer, Trainer,
                          TrainingArguments, DataCollatorForLanguageModeling)
from datasets import load_from_disk

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

BASE = "Qwen/Qwen2.5-Coder-3B-Instruct"
OUT  = "/weka/s225250685/mats-tist/alignment-handbook/output/planner-v1-qwen3b-griffith-sft"
MAX_LEN = 6144

print(f"Loading {BASE}...", flush=True)
tok = AutoTokenizer.from_pretrained(BASE, trust_remote_code=True,
    cache_dir="/weka/s225250685/Huggingface/hub")
if tok.pad_token is None:
    tok.pad_token = tok.eos_token

model = AutoModelForCausalLM.from_pretrained(
    BASE, torch_dtype=torch.bfloat16, trust_remote_code=True,
    attn_implementation="sdpa",
    cache_dir="/weka/s225250685/Huggingface/hub",
)
print("Model loaded.", flush=True)

dd = load_from_disk("data/hf_planner_sft_griffith")
print(f"Dataset C: train={len(dd['train'])} test={len(dd['test'])}", flush=True)

def encode(ex):
    # Qwen chat format: user message = griffith schema + Planning:
    # assistant message = full CoT completion
    prompt = ex["prompt"]      # ends with "Planning:"
    cot    = ex["completion"]  # Goal→Condition→Tables→Final SQL
    text = (f"<|im_start|>user\n{prompt}<|im_end|>\n"
            f"<|im_start|>assistant\n{cot}<|im_end|>")
    return tok(text, truncation=True, max_length=MAX_LEN, padding=False)

train_ds = dd["train"].map(encode, remove_columns=dd["train"].column_names, num_proc=4)
eval_ds  = dd["test"].map(encode,  remove_columns=dd["test"].column_names,  num_proc=4)
collator = DataCollatorForLanguageModeling(tok, mlm=False)

targs = TrainingArguments(
    output_dir=OUT,
    num_train_epochs=3,
    per_device_train_batch_size=1,
    per_device_eval_batch_size=1,
    gradient_accumulation_steps=8,
    learning_rate=2e-5,
    warmup_ratio=0.05,
    lr_scheduler_type="cosine",
    bf16=True,
    logging_steps=10,
    save_strategy="epoch",
    eval_strategy="epoch",
    save_total_limit=1,
    report_to=[],
    gradient_checkpointing=True,
    gradient_checkpointing_kwargs={"use_reentrant": False},
    remove_unused_columns=False,
    dataloader_num_workers=2,
)
trainer = Trainer(
    model=model, args=targs,
    train_dataset=train_ds, eval_dataset=eval_ds,
    tokenizer=tok, data_collator=collator,
)
trainer.train()
trainer.save_model(OUT)
tok.save_pretrained(OUT)
print(f"SAVED: {OUT}", flush=True)
PYEOF

[ -f "$OUT_PLANNER/config.json" ] || { echo "PLANNER SFT FAILED" | tee -a "$LOG"; exit 1; }
echo "==== [A] planner saved: $OUT_PLANNER ====" | tee -a "$LOG"

##############################################
# STAGE B: Oracle comparison K=4 on 200 BIRD-dev questions
#   New Qwen planner vs Qwen2.5-Coder-3B-Instruct baseline (no fine-tune)
##############################################
echo "==== [B] oracle check — new Qwen planner (K=4, 200q) ====" | tee -a "$LOG"

$VLLM serve "$OUT_PLANNER" --served-model-name planner --port 8100 \
    --dtype bfloat16 --gpu-memory-utilization 0.85 \
    --enforce-eager --max-model-len 8192 > "${LOG}.p" 2>&1 &
wait_url http://localhost:8100/v1/models && echo "  new Qwen planner READY" | tee -a "$LOG"

OUT_ORACLE=eval_results/planner_v1_qwen3b_griffith_sft_K4_200q.jsonl
rm -f "$OUT_ORACLE"
$PY scripts/run_pipeline_rollouts.py \
    --input_file data/sft_bird_with_evidence_dev_text2sql.json \
    --output_file "$OUT_ORACLE" \
    --planner_host http://localhost:8100 \
    --planner_format qwen \
    --validator_host none --fixer_host none \
    --K 4 --temperature 1.0 --top_p 0.9 \
    --max_planner_tokens 1024 \
    --max_questions 200 --n_threads 4 2>&1 | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$OUT_ORACLE" qwen3b_griffith_sft_K4 2>&1 | tee -a "$LOG"

kill_vllm

echo "==== [B] baseline: Qwen2.5-Coder-3B-Instruct no fine-tune ====" | tee -a "$LOG"
$VLLM serve "$BASE" --served-model-name planner --port 8100 \
    --dtype bfloat16 --gpu-memory-utilization 0.85 \
    --enforce-eager --max-model-len 8192 > "${LOG}.p2" 2>&1 &
wait_url http://localhost:8100/v1/models && echo "  baseline READY" | tee -a "$LOG"

OUT_BASE=eval_results/planner_qwen3b_base_K4_200q.jsonl
rm -f "$OUT_BASE"
$PY scripts/run_pipeline_rollouts.py \
    --input_file data/sft_bird_with_evidence_dev_text2sql.json \
    --output_file "$OUT_BASE" \
    --planner_host http://localhost:8100 \
    --planner_format qwen \
    --validator_host none --fixer_host none \
    --K 4 --temperature 1.0 --top_p 0.9 \
    --max_planner_tokens 1024 \
    --max_questions 200 --n_threads 4 2>&1 | tee -a "$LOG"
$PY scripts/compute_bestofn_metrics.py "$OUT_BASE" qwen3b_base_K4 2>&1 | tee -a "$LOG"

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