selimaktas/turkish-flow-drafter-prompts
Turkish prompts for Chained-Flow drafter training Chat-templated Turkish prompts used to train and evaluate the Turkish Flow-Drafter checkpoints for Qwen/Qwen3.5-4B / 9B / 27B. Prompts only — no completions. A drafter is trained on the target model's own hidden states, so continuations are generated locally by running the target over these prompts. Nothing here is a model output. split rows what it is v1/ 29,100 train + 300 holdout the mixture the released Turkish… See the full description on the dataset page: https://huggingface.co/datasets/selimaktas/turkish-flow-drafter-prompts.
Turkish prompts for Chained-Flow drafter training
Chat-templated Turkish prompts used to train and evaluate the Turkish Flow-Drafter checkpoints for Qwen/Qwen3.5-4B / 9B / 27B.
Prompts only — no completions. A drafter is trained on the target model's own hidden states, so continuations are generated locally by running the target over these prompts. Nothing here is a model output.
Sources
Prompts are extracted from four public Turkish datasets:
A source's name is not its row count: 3,000 rows of the 20k function-calling set and 900 of the 10k tool-calling set were used. The mixture is sized by token count rather than row count, because prompt lengths differ by 2-5x across these sources — 185 tokens on average for InstrucTurca against 840 for tool-calling — so matching on rows would overshoot the budget several times over. v1 is 7,621,659 prompt tokens; v2 is 19,083,633.
Format
One JSON object per line, {"prompt": str, "prompt_tokens": int}. Prompts are already chat-templated for Qwen3.5 with thinking disabled — the <think>\n\n</think>\n\n prefix is baked into the text. Any in-engine benchmark must send chat_template_kwargs: {"enable_thinking": false} or it will measure a different distribution from the one these were built for.
v2 extends v1 exactly
Same seed; the builder fills the holdout first from a seeded shuffle, so raising the per-source train counts preserves both the holdout and the existing train rows and appends new ones. Verified byte-for-byte. Practically: teacher states collected against v1 stay valid, only the tail needs collecting, and v1-vs-v2 comparisons run on an identical evaluation set.
ood/ is deliberately independent
turkish_alpaca_100 and wikirag_tr_100 come from sources absent from the training mixture, so they measure whether Turkish ability transfers rather than how well a checkpoint fits its own corpus. Their absolute values are not comparable to the in-distribution holdout — prompts average 37 tokens for the alpaca split, so 72 of 100 rows hit the 256-token generation cap — but the ordering between checkpoints is the signal, and that is what they are for.