draft
Datasets
All datasets matching “draft”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.dlm-exp1-drafter-voting-research-artifacts
dlm-exp1 drafter-voting: research artifacts (public mirror)
Everything bulky from the drafter-voting branch of https://github.com/NoviceCoderInfinity/dlm-general-expert-exp1
that git does not track (uploaded 2026-09-06 before the compute instance was destroyed):
logs/research_v1/: baseline cells (one JSON line per item; per-pass NPZ traces for both checkpoints in
baselines_traces/ and baselines_b32_traces/), state-batch fidelity probe states (fidelity/), flush-fusion
boundaries… See the full description on the dataset page: https://huggingface.co/datasets/Arushhh/dlm-exp1-drafter-voting-research-artifacts.dlm-exp1-drafter-voting-research-artifacts
dlm-exp1 drafter-voting research artifacts (public)
Draft-aligned judge LoRA for SDAR-30B-A3B-Chat-b32 with a 1.7B drafter and DES expert pooling: goal = decode speedup with minimal accuracy loss vs the stock model.
model / setting
weights
eval records
stock SDAR-30B-A3B-Chat-b32 (+ 1.7B drafter where used)
JetLM/SDAR-30B-A3B-Chat-b32 @ c351bbc3… (Hugging Face)
reference_evals/da_decode (*_stock), r3_A/evals/stock
stock + DES32 / DES48 (decode-time expert pooling… See the full description on the dataset page: https://huggingface.co/datasets/Anupam-Rawat-IITB/dlm-exp1-drafter-voting-research-artifacts.data_draft
TechJam 2026 Data Draft
Dataset Summary
This private research draft validates a binary image-classifier pipeline:
the local data loader, feature extractor, classifier head, calibration step,
and robustness evaluation. It is not a public benchmark, a claim of real-world
detection quality, or part of the production 80k-master corpus.
The package has 10,000 canonical RGB PNG images. Every image has a
portable manifest record with a label, split, source provenance… See the full description on the dataset page: https://huggingface.co/datasets/Joshyxwa/data_draft.draftfm-frozen-eval
DraftFM frozen_eval: the per-pick prediction archive
Every prediction behind the tables in the DraftFM paper, archived so the
results can be recomputed (no GPU needed) and audited pick by pick.
Model: brianward92/draftfm
Code and paper: https://github.com/brianward92/mtga
Each top-level directory is named by the sha256 of the frozen data snapshot
it was evaluated against — the directory name itself is the registration.
Inside are per-pick prediction parquets for every model in… See the full description on the dataset page: https://huggingface.co/datasets/brianward92/draftfm-frozen-eval.wwi-draft-cards