datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Qwen3.8-27B-Drafter-SFT
Qwen3.8-27B Drafter SFT Corpus
Supervised fine-tuning data released for training speculative drafters for Qwen/Qwen3.8-27B. All completions were generated with Qwen/Qwen3.8-27B at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
The dataset contains 367,535 source conversations and 450,401 train rows, totaling 1,953,218,671 tokens after filtering and evaluation decontamination. Rows contain Qwen3.8-27B-tokenized prompts and target-generated completions, together with loss… See the full description on the dataset page: https://huggingface.co/datasets/DaoCloud/Qwen3.8-27B-Drafter-SFT.turkish-flow-drafter-prompts
GitHub repo ·
Technical blog ·
Model collection
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/… See the full description on the dataset page: https://huggingface.co/datasets/ytu-ce-cosmos/turkish-flow-drafter-prompts.DrafterBench
Dataset Card for DrafterBench
Dataset Description
DrafterBench is a large-scale toolkit focused on evaluating the proficiency of Large Language Models (LLMs) in automating Civil Engineering tasks.
This dataset hosts a task suite summarized across 20 real-world projects, encompassing a total of 1920 tasks.
It replicates the complexity of real-world engineering tasks and provides a technical platform to test the four key capabilities of LLMs:
Structured data understanding… See the full description on the dataset page: https://huggingface.co/datasets/Eason666/DrafterBench.english-flow-drafter-prompts
GitHub repo ·
Technical blog ·
Model collection
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.
split
rows
prompt tokens
what it is
v1/… See the full description on the dataset page: https://huggingface.co/datasets/ytu-ce-cosmos/english-flow-drafter-prompts.p2026-007-qwen38-27b-2bit-drafting-results
Self-Drafting Qwen3.8-27B at Two Bits: Speed, Energy, and Memory on a 12 GB Laptop GPU — result index
Matthew Schwartz — ORCID 0009-0009-4171-7247
This dataset accompanies “Self-Drafting Qwen3.8-27B at Two Bits: Speed, Energy, and Memory on a 12 GB Laptop GPU.” The immutable tagged report and full
artifact use the following versioned path:
https://github.com/matthematics1137/research-artifacts/tree/drafted-2bit-local-efficiency-v1.0.0/papers/2026-drafted-2bit-local-efficiency… See the full description on the dataset page: https://huggingface.co/datasets/mv1137/p2026-007-qwen38-27b-2bit-drafting-results.bonsai2-drafter-eval
Bonsai 2 drafter evaluation corpora
Prompts and greedy responses from PrismML's
Ternary-Bonsai-2-27B, recorded
as token ids together with the per-round accepted lengths of the speculative decoding loop that
produced them. The data exists to evaluate and train DFlash 2 drafters against this one target.
Target
prism-ml/Ternary-Bonsai-2-27B-mlx-2bit (MLX pack), greedy, temperature 0
Loop
mlx-dspark 0.18.0 DFlash 2 loop, stock drafter z-lab/Qwen3.8-27B-DFlash2, draft… See the full description on the dataset page: https://huggingface.co/datasets/Schiltmans/bonsai2-drafter-eval.claim_drafter
Claim Drafter — datasets
Training and evaluation data for vishwr/claim_drafter,
a LoRA adapter that drafts US patent claims from a plain-English invention disclosure.
Two datasets are included:
sft — 9,662 train / 1,314 validation examples in conversational (messages)
format. Targets are as-GRANTED claims fetched per patent number (not the
as-filed claims that ship with HUPD), and prompts are stripped of the patent
summary so the model must draft rather than reformat.
dpo — 5… See the full description on the dataset page: https://huggingface.co/datasets/vishwr/claim_drafter.english-flow-drafter-prompts
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.
split
rows
prompt tokens
what it is
v1/
19,672 train + 500 holdout
1,518,620
the original mixture… See the full description on the dataset page: https://huggingface.co/datasets/selimaktas/english-flow-drafter-prompts.Patient-Message-Response-DraftingPaper: How Much Would a Clinician Edit This Draft? Evaluating LLM Alignment for Patient Message Response Drafting (arxiv link)
Dataset Details:
The patient message response drafting dataset is designed to evaluate how well LLMs respond to patient messages in patient portal communication.
Each semi-synthetic patient message is paired with a real de-identified EHR from a patient at our collaborating hospital.
Each doctor response is written by a clinician, guided by clinician response themes… See the full description on the dataset page: https://huggingface.co/datasets/PortalPal-AI/Patient-Message-Response-Drafting.chrono-2020-leak-sft-llm-draft
ChronoLLM 2020 leak-SFT (LLM draft)
Public gated (gated=manual) prompt/phrase pairs for SN38 leak training
(LLM gen+judge, object-first phrases) over the 2020 harvest.
Final snapshot (stopped):
accepted pairs: 2,222,376
trainable tokens (GPT-2, prompt+phrase): 79,901,922
closed shards on Hub: 1116 across 481 chunks
Repo: jjjlimaus/chrono-2020-leak-sft-llm-draft
Layout
chunk-YYYYMMDDHHMM/
COMPLETE.json
llm/part-NNNNNN.jsonl.gz
Trainers must read chunk-* in… See the full description on the dataset page: https://huggingface.co/datasets/jjjlimaus/chrono-2020-leak-sft-llm-draft.DrafterBench
Dataset Card for DrafterBench
DrafterBench
DrafterBench is a large-scale toolkit focused on evaluating the proficiency of Large Language Models (LLMs) in automating Civil Engineering tasks.
The dataset contains tasks derived from real-world engineering drawing revision processes.
This dataset is released for anonymous review.
Code: https://github.com/anonymous733882/DrafterBench
This dataset hosts a task suite summarized across 20 real-world projects, encompassing a total… See the full description on the dataset page: https://huggingface.co/datasets/anonymous733882/DrafterBench.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.opengloss-v1.1-drafting
OpenGloss Drafting v1.1
Dataset Summary
OpenGloss Drafting is a synthetic dataset of educational content drafts generated from
vocabulary terms and encyclopedia entries. Each draft is a self-contained piece of writing
(article, story, memo, essay, etc.) that naturally incorporates specific vocabulary terms
with their definitions and context.
This dataset supports curriculum-aligned content generation, vocabulary-in-context learning,
and educational text synthesis. It is… See the full description on the dataset page: https://huggingface.co/datasets/mjbommar/opengloss-v1.1-drafting.mismatched-wrong-drafts
Per-variant training data
Training data for "Weak-to-Strong Elicitation via Mismatched Wrong Drafts" (Wei Deng, 2026).
Code: https://github.com/weiddeng/mismatched-wrong-drafts
Models: the four trained models below
Paper: https://arxiv.org/abs/2605.17314
Each row/training datapoint is a MATH problem with a draft injected into the prompt
field.
Variant
Draft shown to the learner
Trained model
# Rows
mismatched_wrong ⭐
a wrong draft from a different problem… See the full description on the dataset page: https://huggingface.co/datasets/hugruby/mismatched-wrong-drafts.DraftNEPABench
DraftNEPABench: A Benchmark for Drafting NEPA Document Sections with Coding Agents
Dataset Description
DraftNEPABench is a novel benchmark designed to evaluate the capabilities of Large Language Models (LLMs) and LLM-based agents in drafting sections of Environmental Impact Statements (EIS). This benchmark is curated by subject matter experts (SMEs) to ensure that the tasks reflect realistic, domain-relevant drafting challenges encountered in environmental planning and… See the full description on the dataset page: https://huggingface.co/datasets/PNNL/DraftNEPABench.