--- language: - id license: other task_categories: - text-generation - question-answering size_categories: - n<1K 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: nvidia_nemotron_3_ultra_550b_a55b_pajak_train data_files: - split: train path: data/nvidia-nemotron-3-ultra-550b-a55b-pajak-train.json --- # SQL-SFT-V1 Dataset distilasi tentang pelaporan pajak di Indonesia dalam Bahasa Indonesia, dibuat untuk belajar membangun LLM sendiri. Formatnya mengikuti [rasbt/math_distill](https://huggingface.co/datasets/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=True, ONLY_CORRECT=True, REQUIRE_THINKING=True. Akurasi teacher sebelum filter: | Model teacher | Sampel | Benar | Dengan thinking | |---|---|---|---| | `nvidia/nemotron-3-ultra-550b-a55b` | 24 | 3 (12.5%) | 24 | ## Files - [data/pajak-eval.json](data/pajak-eval.json): * 1 soal evaluasi (held-out, tidak ada di file train) beserta jawaban teacher. - [data/pajak-train.json](data/pajak-train.json): * 2 sampel train dari seluruh teacher (tidak tumpang tindih dengan set eval). - [data/pajak-train_4000.json](data/pajak-train_4000.json): * Seperti pajak-train.json tetapi hanya sampel dengan panjang <= 4000 token (setelah chat template): 2 sampel. - [data/nvidia-nemotron-3-ultra-550b-a55b-pajak-train.json](data/nvidia-nemotron-3-ultra-550b-a55b-pajak-train.json): * Sampel train yang dijawab oleh model `nvidia/nemotron-3-ultra-550b-a55b`: 2 sampel. ## Fields Each example contains: - `problem`: soal perpajakan - `gtruth_answer`: jawaban singkat (kunci) - `message_thinking`: thinking trace dari teacher - `message_content`: jawaban akhir dari teacher Keduanya dapat digabung, sama seperti math_distill: ```python complete_answer = f"{data['message_thinking']}\n\n{data['message_content']}" ``` ## Dataset Statistics Dihitung dengan `scripts/average_field_lengths_json.py`, misalnya: ```bash python scripts/average_field_lengths_json.py --json_path data/pajak-eval.json --tokenizer Qwen/Qwen3-0.6B ``` ### `data/pajak-eval.json` ``` Records: 1 Tokenizer: Qwen/Qwen3-0.6B Field AvgTokens MinTokens MaxToken Count gtruth_answer 240.00 240 240 1 message_content 231.00 231 231 1 message_thinking 624.00 624 624 1 problem 112.00 112 112 1 Verified-correct answers: 1/1 (100.0%) ``` ### `data/pajak-train.json` ``` Records: 2 Tokenizer: Qwen/Qwen3-0.6B Field AvgTokens MinTokens MaxToken Count gtruth_answer 146.00 134 158 2 message_content 78.50 72 85 2 message_thinking 556.00 155 957 2 problem 123.00 115 131 2 Verified-correct answers: 2/2 (100.0%) ``` ### `data/pajak-train_4000.json` ``` Records: 2 Tokenizer: Qwen/Qwen3-0.6B Field AvgTokens MinTokens MaxToken Count gtruth_answer 146.00 134 158 2 message_content 78.50 72 85 2 message_thinking 556.00 155 957 2 problem 123.00 115 131 2 Verified-correct answers: 2/2 (100.0%) ``` ### `data/nvidia-nemotron-3-ultra-550b-a55b-pajak-train.json` ``` Records: 2 Tokenizer: Qwen/Qwen3-0.6B Field AvgTokens MinTokens MaxToken Count gtruth_answer 146.00 134 158 2 message_content 78.50 72 85 2 message_thinking 556.00 155 957 2 problem 123.00 115 131 2 Verified-correct answers: 2/2 (100.0%) ``` ## Intended use Untuk pembelajaran pribadi: melatih/distilasi LLM kecil. ## Usage ```python from datasets import load_dataset repo_id = "bloksbgt/SQL-SFT-V1" 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()) ```