CoolFace
Datasetpublic

CryptoYogi/vazhi-tamil-sft-v5_0

VAZHI Tamil SFT Dataset v5.0 Tamil-first SFT dataset for the VAZHI project - a free, offline Tamil AI assistant. Dataset Stats Split Samples Train 5,328 Eval 593 Total 5,921 Bucket Distribution Source Count % vazhi_packs 2,961 50.0% safety 1,800 30.4% sadhguru_qa 848 14.3% thirukkural 168 2.8% handcrafted 120 2.0% general 24 0.4% Key Improvements over v4.1 100% Tamil responses (v4.1 had… See the full description on the dataset page: https://huggingface.co/datasets/CryptoYogi/vazhi-tamil-sft-v5_0.

sourceHugging Facecc-by-sa-4.0updated 7mo agoView on Hugging Face
0likes5downloads
Dataset Card

VAZHI Tamil SFT Dataset v5.0

Tamil-first SFT dataset for the VAZHI project - a free, offline Tamil AI assistant.

Dataset Stats

SplitSamples
Train5,328
Eval593
Total5,921

Bucket Distribution

SourceCount%
vazhi_packs2,96150.0%
safety1,80030.4%
sadhguru_qa84814.3%
thirukkural1682.8%
handcrafted1202.0%
general240.4%

Key Improvements over v4.1

  • —100% Tamil responses (v4.1 had 75.8% garbage English)
  • —Short responses: avg 41 words (v4.1 had 200-1000 word MT essays)
  • —Authentic Tamil sources: Sadhguru articles restructured into Q&A (not machine-translated)
  • —Domain-relevant: Tamil Nadu focused content
  • —Clean 5.9K samples (v4.1 had 13K padded with garbage)
  • —Thirukkural in Q&A format only (no verbatim recitation that could cause parroting)

Data Sources

  1. 1.Vazhi-packs v5 (2,961): 6 domain packs (healthcare, education, legal, govt, security, culture) with Tamil responses
  2. 2.IndicAlign safety (1,800): Tamil safety refusals from HHRLHFT + ToxicMatrix
  3. 3.Sadhguru Q&A (848): High-quality Tamil Q&A restructured from isha.sadhguru.org Tamil articles
  4. 4.Thirukkural Q&A (168): Interpretive Q&A about Thirukkural verses (no verbatim)
  5. 5.Handcrafted (120): Product voice - greetings, refusals, brevity, guardrails
  6. 6.General (24): Numbers, weather, health basics

Format

ChatML format with Tamil system prompt. Each sample has:

  • —text: Full ChatML-formatted conversation
  • —bucket: Source category
  • —source: Original data source
  • —category: Content category

Quality Metrics

  • —Tamil char % avg: 85.2%
  • —Word count avg: 41 words
  • —Word count range: 2-142 words