19. 대규모 학습 데이터 수집

pip install datasets huggingface_hub pyarrow
mkdir -p data/math/sft data/math/dpo data/finance data/reasoning

python -c "
from datasets import load_dataset
import json

ds = load_dataset('microsoft/orca-math-word-problems-200k', split='train')
count = 0
with open('data/math/sft/math_orca_200000.jsonl', 'w', encoding='utf-8') as f:
    for row in ds:
        q = row.get('question') or row.get('instruction') or row.get('problem')
        a = row.get('answer') or row.get('output') or row.get('response')
        if q and a:
            f.write(json.dumps({
                'prompt': f'Solve the following math problem step by step.\n\nProblem: {q.strip()}',
                'response': a.strip(),
            }, ensure_ascii=False) + '\n')
            count += 1
print(f'Orca-Math SFT: {count} rows saved -> data/math/sft/math_orca_200000.jsonl')
"

head -35 data/math/sft/math_orca_200000.jsonl > data/math/sft/math_orca_35.jsonl
echo "Bench: $(wc -l < data/math/sft/math_orca_35.jsonl) rows"

python -c "
from datasets import load_dataset
import json

ds = load_dataset('xinlai/Math-Step-DPO-10K', split='train')
count = 0
with open('data/math/dpo/math_step_dpo_10700.jsonl', 'w', encoding='utf-8') as f:
    for row in ds:
        prompt = row.get('prompt', '')
        chosen = row.get('full_chosen') or row.get('chosen', '')
        rejected = row.get('full_rejected') or row.get('rejected', '')
        if prompt and chosen and rejected:
            f.write(json.dumps({
                'prompt': prompt.strip(),
                'chosen': chosen.strip(),
                'rejected': rejected.strip(),
                'dataset': row.get('dataset', ''),
                'answer': str(row.get('answer', '')),
            }, ensure_ascii=False) + '\n')
            count += 1
print(f'Math-Step-DPO: {count} rows saved -> data/math/dpo/math_step_dpo_10700.jsonl')
"

# DPO 벤치에서는 SFT 벤치 데이터를 재사용 (math_orca_35.jsonl)
ls -la data/math/sft/math_orca_35.jsonl

python -c "
import json, ast
from pathlib import Path
from datasets import load_dataset

OUT = Path('data/finance')
OUT.mkdir(parents=True, exist_ok=True)

def clean(x):
    if x is None: return ''
    return str(x).strip()

def write_jsonl(path, rows):
    with open(path, 'w', encoding='utf-8') as f:
        for r in rows:
            f.write(json.dumps(r, ensure_ascii=False) + '\n')
    return len(rows)

all_rows = []

# 1) Won-Instruct (KRX)
print('[1/5] Won-Instruct...')
ds = load_dataset('KRX-Data/Won-Instruct', split='train')
for row in ds:
    p = clean(row.get('prompt'))
    r = clean(row.get('original_response') or row.get('response'))
    if p and r:
        all_rows.append({'prompt': p, 'response': r})
print(f'  Won-Instruct: {len(all_rows)} rows')

# 2) financial-mmlu-ko
print('[2/5] financial-mmlu-ko...')
n0 = len(all_rows)
ds = load_dataset('allganize/financial-mmlu-ko', split='test')
for row in ds:
    msgs = row.get('messages', [])
    if not isinstance(msgs, list) or len(msgs) < 2: continue
    parts_u, parts_a = [], []
    for m in msgs:
        role = clean(m.get('role','')).lower()
        content = clean(m.get('content',''))
        if role in ('user','human') and content: parts_u.append(content)
        elif role in ('assistant','gpt') and content: parts_a.append(content)
    if parts_u and parts_a:
        all_rows.append({'prompt': '\n\n'.join(parts_u), 'response': '\n\n'.join(parts_a)})
print(f'  financial-mmlu-ko: {len(all_rows)-n0} rows')

# 3) flare-convfinqa-ko
print('[3/5] flare-convfinqa-ko...')
n0 = len(all_rows)
ds = load_dataset('allganize/flare-convfinqa-ko', split='test')
for row in ds:
    msgs = row.get('messages', [])
    if not isinstance(msgs, list) or len(msgs) < 2: continue
    parts_u, parts_a = [], []
    for m in msgs:
        role = clean(m.get('role','')).lower()
        content = clean(m.get('content',''))
        if role in ('user','human') and content: parts_u.append(content)
        elif role in ('assistant','gpt') and content: parts_a.append(content)
    if parts_u and parts_a:
        all_rows.append({'prompt': '\n\n'.join(parts_u), 'response': '\n\n'.join(parts_a)})
print(f'  flare-convfinqa-ko: {len(all_rows)-n0} rows')

# 4) KorfinQA
print('[4/5] KorfinQA...')
n0 = len(all_rows)
ds = load_dataset('mssongit/KorfinQA', split='train')
for row in ds:
    q = clean(row.get('question'))
    a = clean(row.get('answer'))
    if q and a:
        all_rows.append({'prompt': q, 'response': a})
print(f'  KorfinQA: {len(all_rows)-n0} rows')

# 5) QA_Instruction (6 configs)
print('[5/5] QA_Instruction (6 configs)...')
n0 = len(all_rows)
for cfg in ['Multiple-Choice QA','Binary QA','Extractive QA',
            'Numerical Reasoning Arithmetic','Numerical Reasoning Comparison',
            'Numerical Reasoning Extraction']:
    try:
        ds = load_dataset('FINNUMBER/QA_Instruction', cfg, split='train')
        for row in ds:
            q = clean(row.get('Q'))
            a = clean(row.get('A'))
            c = clean(row.get('C'))
            if q and a:
                prompt = f'[{cfg}]\n\n[문맥]\n{c}\n\n[질문]\n{q}' if c else q
                all_rows.append({'prompt': prompt, 'response': a})
    except Exception as e:
        print(f'  {cfg}: skipped ({e})')
print(f'  QA_Instruction: {len(all_rows)-n0} rows')

# 저장
n = write_jsonl(OUT / 'finance_ko_sft_full.jsonl', all_rows)
print(f'\nTotal: {n} rows -> data/finance/finance_ko_sft_full.jsonl')
"

# 전체에서 마지막 1000줄을 벤치로, 나머지를 훈련으로
total=$(wc -l < data/finance/finance_ko_sft_full.jsonl)
train=$((total - 1000))
head -${train} data/finance/finance_ko_sft_full.jsonl > data/finance/finance_ko_sft_179k_raw.jsonl
tail -1000 data/finance/finance_ko_sft_full.jsonl > data/finance/finance_ko_sft_1k_raw.jsonl

echo "Train: $(wc -l < data/finance/finance_ko_sft_179k_raw.jsonl) rows"
echo "Bench: $(wc -l < data/finance/finance_ko_sft_1k_raw.jsonl) rows"

python -c "
import json
from huggingface_hub import HfApi, hf_hub_download
import pyarrow.parquet as pq

repo_id = 'open-r1/OpenR1-Math-220k'
api = HfApi()
files = api.list_repo_files(repo_id=repo_id, repo_type='dataset')
parquet_files = sorted([f for f in files if f.endswith('.parquet')
                        and not f.startswith('extended/')])

print(f'Found {len(parquet_files)} parquet files')

MAX_ROWS = 41000  # 40k train + 1k bench
written = 0

with open('data/reasoning/openr1_math_full.jsonl', 'w', encoding='utf-8') as fout:
    for repo_file in parquet_files:
        local_path = hf_hub_download(repo_id=repo_id, repo_type='dataset', filename=repo_file)
        pf = pq.ParquetFile(local_path)
        for batch in pf.iter_batches(batch_size=2048):
            for row in batch.to_pylist():
                prompt = (row.get('problem') or '').strip()
                response = (row.get('solution') or '').strip()
                if prompt and response:
                    fout.write(json.dumps({'prompt': prompt, 'response': response},
                                         ensure_ascii=False) + '\n')
                    written += 1
                    if written >= MAX_ROWS:
                        break
            if written >= MAX_ROWS:
                break
        if written >= MAX_ROWS:
            break

print(f'OpenR1-Math: {written} rows -> data/reasoning/openr1_math_full.jsonl')
"

head -40000 data/reasoning/openr1_math_full.jsonl > data/reasoning/openr1_math_40k.jsonl
tail -1000 data/reasoning/openr1_math_full.jsonl > data/reasoning/openr1_math_1k.jsonl

echo "Train: $(wc -l < data/reasoning/openr1_math_40k.jsonl) rows"
echo "Bench: $(wc -l < data/reasoning/openr1_math_1k.jsonl) rows"

data/
├── math/
│   ├── sft/
│   │   ├── math_orca_200000.jsonl    # Orca-Math SFT (200K)
│   │   └── math_orca_35.jsonl        # 벤치용 (35 rows)
│   └── dpo/
│       └── math_step_dpo_10700.jsonl # Math-Step DPO (10.7K)
├── finance/
│   ├── finance_ko_sft_179k_raw.jsonl # 훈련용 (~179K)
│   └── finance_ko_sft_1k_raw.jsonl   # 벤치용 (1K)
├── reasoning/
│   ├── openr1_math_40k.jsonl         # CoT 훈련용 (40K)
│   └── openr1_math_1k.jsonl          # CoT 벤치용 (1K)
├── sft_50k_en_ko_raw.jsonl           # 일반 SFT (기본 제공)
└── sft_1k_en_ko_raw.jsonl            # 일반 SFT 벤치 (기본 제공)