kumar2235/kannada-sft-instruction-pool

Kannada SFT Instruction Pool

A 32,888-example synthetic instruction-tuning dataset for Kannada conversational language models. It was generated primarily in Kannada with Gemma 3 27B and organized into five pools. It is intended for supervised fine-tuning (SFT) and experimentation with Kannada assistant behavior, grounded responses, code-mixed requests, and multi-turn conversations.

Intended use

Use this dataset to train or evaluate Kannada chat assistants, especially systems that need to:

• respond to everyday questions, advice requests, explanations, rewrites, letters, and small talk;
• answer or summarize using a supplied Kannada passage;
• handle multi-turn follow-ups, corrections, preferences, and shortening requests;
• respond to Kannada-English code-mixed requests and simple programming tasks.

This is text-only chat data. It does not contain audio, ASR labels, or a speech dataset. It is not a standalone factual knowledge base or a benchmark with independently verified answers. Validate model behavior on separate human-reviewed evaluations before relying on it in production.

Dataset composition and splits

The final data is split by meta.base_id, so examples derived from the same base topic are kept in the same split. The project report records no train/validation base-group overlap.

Pool	Train	Validation	Total	Share
Behavior	11,477	204	11,681	35.5%
Grounded	7,627	331	7,958	24.2%
Multi-turn	6,540	244	6,784	20.6%
Code-mixed	4,940	223	5,163	15.7%
Facts	1,247	55	1,302	4.0%
Total	31,831	1,057	32,888	100%

The validation split is for development-time evaluation and model selection. It is not an independent benchmark: it comes from the same generation pipeline and should not replace a separately collected evaluation set. The source build report targeted 30,000 examples; the finalized split contains 32,888.

Pool descriptions

• Behavior: small talk, advice, rewriting, letter writing, explanations, and a small number of translation tasks.
• Grounded: question answering and summaries conditioned on a supplied passage.
• Multi-turn: follow-up conversations, naming, preference, correction, and shortening tasks.
• Code-mixed: Kannada-English requests and programming-related tasks.
• Facts: short answers constrained to supplied source passages; some rows overlap in task type with grounded QA, but are separately labeled by pool.

Data format

The files are JSON Lines. Each row contains:

• messages: an ordered list of { "role", "content" } chat turns. Roles alternate between user and assistant; records contain between one and four assistant turns.
• meta: provenance fields including task_id, pool, kind, base_id, and teacher; source-backed rows also retain the available source ID, URL, title, attribution, and license label.

Example:

{"messages":[{"role":"user","content":"ಬೇಸರವಾಗುತ್ತಿದೆ, ಏನಾದರೂ ಮಾತಾಡೋಣ."},{"role":"assistant","content":"ಖಂಡಿತ, ಏನ ಬಗ್ಗೆ ಮಾತನಾಡಬೇಕು? ನಿಮ್ಮ ಆಸಕ್ತಿಯ ವಿಷಯಗಳನ್ನು ಹೇಳಿ."}],"meta":{"task_id":"e2e96941358c7f99","pool":"behavior","kind":"smalltalk","base_id":"smalltalk:st11","teacher":"gemma3:27b"}}

Load from the Hub with ?? Datasets:

from datasets import load_dataset

ds = load_dataset("kumar2235/kannada-sft-instruction-pool")
print(ds["train"][0]["messages"])

License and attribution

The uploader applies CC BY-SA 4.0 to this dataset release. Reusers must follow the license terms, including attribution, indicating changes, and ShareAlike obligations for adaptations. License text.

The source artifacts identify the passage catalog as wikimedia/wikipedia:20231101.kn, attribute it to Wikimedia contributors, and record the source license as CC BY-SA (the source report does not specify a version). The release preserves per-example source metadata where it was present in the generation log: 9,247 rows have source/license metadata, including source IDs and URLs; 23,641 rows have no source license recorded. Absence of a source record does not establish that an example has no external influence. Check each row's meta when reusing source-backed material, retain its attribution and URL, and indicate that the passage was transformed into a generated chat example. The dataset-level license does not erase third-party rights in source material.

The teacher tag identifies a Gemma-family model. See Google's Gemma Terms of Use for terms relevant to Gemma; this dataset card does not claim that those terms replace the dataset license or third-party source terms.