SQL-SFT-V1-all-rows / README.md
bloksbgt's picture
Upload SQL-SFT-V1-all-rows
bd4d1ad verified
|
Raw History Blame Contribute Delete
10.9 kB
metadata
language:
  - id
license: other
task_categories:
  - text-generation
  - question-answering
size_categories:
  - 1K<n<10K
configs:
  - config_name: pajak_eval
    data_files:
      - split: train
        path: data/pajak-eval.json
  - config_name: pajak_train
    data_files:
      - split: train
        path: data/pajak-train.json
  - config_name: pajak_train_4000
    data_files:
      - split: train
        path: data/pajak-train_4000.json
  - config_name: deepseek_v4_1_flash_1017_pajak_train
    data_files:
      - split: train
        path: data/deepseek-v4-1-flash-1017-pajak-train.json
  - config_name: deepseek_v4_1_flash_pajak_train
    data_files:
      - split: train
        path: data/deepseek-v4.1-flash-pajak-train.json
  - config_name: deepseek_v4_1_flash_free_pajak_train
    data_files:
      - split: train
        path: data/deepseek-v4.1-flash-free-pajak-train.json
  - config_name: gemini_3_5_flash_lite_pajak_train
    data_files:
      - split: train
        path: data/gemini-3.5-flash-lite-pajak-train.json
  - config_name: gpt_6_luna_pajak_train
    data_files:
      - split: train
        path: data/gpt-6-luna-pajak-train.json
  - config_name: mm_MiniMax_M3_1_Flash_pajak_train
    data_files:
      - split: train
        path: data/mm-MiniMax-M3.1-Flash-pajak-train.json
  - config_name: nvidia_nemotron_3_ultra_550b_a55b_pajak_train
    data_files:
      - split: train
        path: data/nvidia-nemotron-3-ultra-550b-a55b-pajak-train.json
  - config_name: nvidia_nemotron_3_ultra_550b_a55b_free_pajak_train
    data_files:
      - split: train
        path: data/nvidia-nemotron-3-ultra-550b-a55b-free-pajak-train.json
  - config_name: z_ai_glm_5_3_flash_pajak_train
    data_files:
      - split: train
        path: data/z-ai-glm-5.3-flash-pajak-train.json

SQL-SFT-V1-all-rows

Dataset distilasi pertanyaan dan jawaban SQL. Formatnya mengikuti rasbt/math_distill.

Construction

  • Soal disediakan sendiri oleh pemilik dataset (file problems.jsonl) beserta kunci jawaban singkat (gtruth_answer).
  • Jawaban (thinking + jawaban akhir) dihasilkan oleh model teacher lewat router API; verifikasi otomatis terhadap kunci jawaban: VERIFY=False, ONLY_CORRECT=False, REQUIRE_THINKING=False.

Akurasi teacher sebelum filter:

Model teacher Sampel Benar Dengan thinking
deepseek-v4-1-flash-1017 224 224 (100.0%) 179
deepseek-v4.1-flash 12 12 (100.0%) 12
deepseek-v4.1-flash-free 40 40 (100.0%) 40
gemini-3.5-flash-lite 3 3 (100.0%) 3
gpt-6-luna 2715 2715 (100.0%) 0
mm/MiniMax-M3.1-Flash 18 18 (100.0%) 18
nvidia/nemotron-3-ultra-550b-a55b 2890 2890 (100.0%) 2890
nvidia/nemotron-3-ultra-550b-a55b:free 64 64 (100.0%) 64
z-ai/glm-5.3-flash 1591 1591 (100.0%) 0

Files

Fields

Each example contains:

  • problem: soal sql
  • gtruth_answer: jawaban singkat (kunci)
  • message_thinking: thinking trace dari teacher
  • message_content: jawaban akhir dari teacher

Keduanya dapat digabung, sama seperti math_distill:

complete_answer = f"<think>{data['message_thinking']}</think>\n\n{data['message_content']}"

Dataset Statistics

Dihitung dengan scripts/average_field_lengths_json.py, misalnya:

python scripts/average_field_lengths_json.py --json_path data/pajak-eval.json --tokenizer Qwen/Qwen3-0.6B

data/pajak-eval.json

Records: 258
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        222.24        173       301    258
message_content      140.90         26       563    258
message_thinking     103.42          0       604    258
problem              138.00        138       138    258

Verified-correct answers: 258/258 (100.0%)

data/pajak-train.json

Records: 7299
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        294.72         44       565   7299
message_content      157.63          4       782   7299
message_thinking     112.90          0     10402   7299
problem              138.15        122       158   7299

Verified-correct answers: 7299/7299 (100.0%)

data/pajak-train_4000.json

Records: 7296
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        294.65         44       565   7296
message_content      157.56          4       782   7296
message_thinking     109.93          0      3254   7296
problem              138.15        122       158   7296

Verified-correct answers: 7296/7296 (100.0%)

data/deepseek-v4-1-flash-1017-pajak-train.json

Records: 214
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        291.47         48       555    214
message_content      225.34         34       776    214
message_thinking     308.66          0      2082    214
problem              137.29        122       158    214

Verified-correct answers: 214/214 (100.0%)

data/deepseek-v4.1-flash-pajak-train.json

Records: 12
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        327.83         79       542     12
message_content      243.67         41       460     12
message_thinking     325.58        213       425     12
problem              139.50        122       153     12

Verified-correct answers: 12/12 (100.0%)

data/deepseek-v4.1-flash-free-pajak-train.json

Records: 40
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        319.05         75       560     40
message_content      465.27        214       762     40
message_thinking     331.85         54      1015     40
problem              141.10        122       158     40

Verified-correct answers: 40/40 (100.0%)

data/gemini-3.5-flash-lite-pajak-train.json

Records: 3
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        321.33        267       404      3
message_content      106.33         57       132      3
message_thinking     284.00        228       356      3
problem              137.67        126       154      3

Verified-correct answers: 3/3 (100.0%)

data/gpt-6-luna-pajak-train.json

Records: 2631
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        297.56         44       565   2631
message_content      120.95         25       434   2631
message_thinking       0.00          0         0   2631
problem              138.06        122       158   2631

Verified-correct answers: 2631/2631 (100.0%)

data/mm-MiniMax-M3.1-Flash-pajak-train.json

Records: 18
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        300.39         75       498     18
message_content      176.22         58       334     18
message_thinking    2336.00        230     10402     18
problem              134.56        122       158     18

Verified-correct answers: 18/18 (100.0%)

data/nvidia-nemotron-3-ultra-550b-a55b-pajak-train.json

Records: 2788
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        293.90         44       562   2788
message_content      135.11          4       705   2788
message_thinking     244.11         29      1361   2788
problem              138.30        122       158   2788

Verified-correct answers: 2788/2788 (100.0%)

data/nvidia-nemotron-3-ultra-550b-a55b-free-pajak-train.json

Records: 61
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        305.89         44       563     61
message_content      137.59         14       414     61
message_thinking     284.03         43       992     61
problem              138.56        122       158     61

Verified-correct answers: 61/61 (100.0%)

data/z-ai-glm-5.3-flash-pajak-train.json

Records: 1532
Tokenizer: Qwen/Qwen3-0.6B
Field             AvgTokens  MinTokens  MaxToken  Count
gtruth_answer        290.34         44       564   1532
message_content      244.11         40       782   1532
message_thinking       0.00          0         0   1532
problem              138.10        122       158   1532

Verified-correct answers: 1532/1532 (100.0%)

Intended use

Untuk pembelajaran pribadi: melatih/distilasi LLM kecil.

Usage

from datasets import load_dataset

repo_id = "bloksbgt/SQL-SFT-V1-all-rows"

train = load_dataset(repo_id, "pajak_train", split="train")
eval_set = load_dataset(repo_id, "pajak_eval", split="train")
train_4000 = load_dataset(repo_id, "pajak_train_4000", split="train")

print(len(train), len(eval_set), len(train_4000))
print(train[0].keys())