ytu-ce-cosmos/english-flow-drafter-prompts

<p align="center"> <img src="assets/chainflow-headline.png" alt="Chain-Flow" width="90%" style="max-width:100%; height:auto;" /> </p>

<p align="center"> <a href="https://github.com/Cosmos-YTU/chain-flow">GitHub repo</a> · <a href="https://chainflow.cosmosytu.com/">Technical blog</a> · <a href="https://huggingface.co/collections/ytu-ce-cosmos/chain-flow">Model collection</a> </p>

English prompts for Chained-Flow drafter training

Chat-templated English prompts used to train the English 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.

The v2 training mixture is `v1` + `v2`, not `v2` alone. v2 widened the mixture by adding five new domains rather than by adding rows to the existing five, so the two directories are disjoint and are kept separate here instead of duplicating v1 inside v2. Together: 33,777 prompts, 4,737,781 prompt tokens.

Sources

Every prompt is rendered from a public dataset with the repo's own formatter, so these files are a replay of what was collected rather than a re-derivation of it.

v1/ — the original mixture

v2/ — the domains v2 added

Summarization is 44% of the combined token count off 8% of the rows — prompt lengths run from 44 to 782 tokens across these sources, so row counts are a poor guide to a source's weight.

A note on nemotron_stem

Its row count is low for a reason worth stating. The collection consumed the first 5,000 rows of the stem split, but those 5,000 rows contain only 669 distinct prompts — 87% are duplicates, and the collector does not deduplicate. These files do deduplicate, so nemotron_stem appears here as 569 train + 100 holdout rather than 5,000. The published rows are the true prompt diversity that source contributed; the collected run simply saw many of them repeatedly.

Format

One JSON object per line, {"prompt": str, "prompt_tokens": int}. Prompts are already chat-templated for Qwen3.5 with enable_thinking=False, so the <think>\n\n</think>\n\n assistant prefix is baked into the text — feed them verbatim, do not re-template. prompt_tokens is the length under the Qwen/Qwen3.5-27B tokenizer.

Holdout rows are taken from the end of each source's range, so the train files are exactly the prompts the collection consumed, in order.

Reproducing

python scripts/build_english_prompts.py --out bench_data_en

from Chained-Flow. It reads the real collect configs for each source's dataset, split, range and formatter, so changing a formatter changes this output.