CoolFace
Datasetpublic

AksaraLLM/aksara-sft-clean-v6

AksaraLLM SFT Clean v6 High-quality Indonesian SFT dataset distilled from Gemini 2.5 Flash Lite via Google Vertex AI, with strict quality gates. Stats Train: 16,752 items Validation: 1,098 items Total: 17,850 items Teacher model: gemini-2.5-flash-lite Task distribution Task type Count factual_qa 4,023 creative 4,003 cultural 3,927 reasoning 3,258 how_to 2,639 Method Curated ~100 Indonesian topic seeds across… See the full description on the dataset page: https://huggingface.co/datasets/AksaraLLM/aksara-sft-clean-v6.

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face
0likes29downloads
Dataset Card

AksaraLLM SFT Clean v6

High-quality Indonesian SFT dataset distilled from Gemini 2.5 Flash Lite via Google Vertex AI, with strict quality gates.

Stats

  • —Train: 16,752 items
  • —Validation: 1,098 items
  • —Total: 17,850 items
  • —Teacher model: gemini-2.5-flash-lite

Task distribution

Task typeCount
factual_qa4,023
creative4,003
cultural3,927
reasoning3,258
how_to2,639

Method

  1. 1.Curated ~100 Indonesian topic seeds across history, geography, culture, food, science, tech, economics, politics, religion, education, law, and daily life.
  2. 2.5 task templates: factualqa, howto, reasoning, creative, cultural.
  3. 3.For each seed × task, prompt Gemini 2.5 Flash Lite (thinkingbudget=0, temperature=0.9, topp=0.95, maxoutputtokens=1800).
  4. 4.Quality gates per item:
  5. 5.Valid JSON with instruction + output
  6. 6.Instruction ≥ 15 chars, output 100–5000 chars
  7. 7.Output has ≥ 50% Latin alphabet characters (weeds out garbage)
  8. 8.No AI/Gemini self-reference phrases ("sebagai model bahasa…", "I am Gemini", etc.)
  9. 9.Deduplication: MD5 on instruction string across the whole set.
  10. 10.Train/val split: deterministic — seed_id starting with 0 → val.

Fields

  • —instruction (str) — user prompt
  • —output (str) — teacher-generated response
  • —task_type (str) — one of {factualqa, howto, reasoning, creative, cultural}
  • —topic (str) — seed topic
  • —source (str) — teacher model identifier
  • —seed_id (str) — reproducibility hash
  • —teacher_reasoning (str) — why the teacher judged the item valuable

Comparison vs v1

DatasetItemsQuality concerns
aksara-sft-clean-v1 (human-derived)33,593inherits remaining noise from source corpora
aksara-sft-clean-v6 (this)17,850controlled distribution, no hallucinated fake-words, no factual blocklist violations

Recommended: mix v1 + v6 for SFT training (v1 gives register variety, v6 gives topic breadth + stylistic quality).

License

Apache 2.0. Outputs are synthetic content from Gemini. Per Google's generative AI terms, synthetic outputs are owned by the user; no attribution required for redistribution.