datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ablation_nemotron_no_thinking
Dataset: ablation_nemotron_no_thinking
This dataset was uploaded from /mnt/yulan_pretrain/mount/data_final_train/ablation_nemotron_no_thinking/stage_1/tmp.
commoncorpus_en_code_fw_ablationkorean-embedding-performance-v1-ablation-200k
Korean Embedding Performance v1 — Ablation 200K
Qwen3-Embedding-8B의 한국어 retrieval continued fine-tuning에서 LoRA/DoRA/부분 및
full fine-tuning, loss, hard-negative 전략을 비교하기 위한 200,000-row 연구·비상업
성능 데이터다. release_eligible: false이며 통합 라이선스는 other다. upstream
source별 조건을 재허가하지 않는다.
구성
계열
Rows
역할
nlpai-lab/ko-triplet-v1.0@1f5d72d
100,254
넓은 한국어 QA/retrieval core
F2 Korean QA/instruction
68,000
webfaq, mqa, koalpaca, realQA, komagpie
F2 retrieval task… See the full description on the dataset page: https://huggingface.co/datasets/LLM-OS-Models/korean-embedding-performance-v1-ablation-200k.screenshot-training-naive-top2-hn-ablation
Chrisyichuan/screenshot-training-naive-top2-hn-ablation
Ablation variant of
Chrisyichuan/screenshot-training-natural-filtered-v2.
Same queries, same positives. Only neg_chunk_paths differ.
The filtered-v2 dataset applies a Gemini VLM judge to filter false negatives
out of the retrieved candidates. This ablation set skips that filter entirely:
for every (query, chunk_path), we hit the text-retrieval search API for the
top-10 results and keep the first two non-positive hits as… See the full description on the dataset page: https://huggingface.co/datasets/Chrisyichuan/screenshot-training-naive-top2-hn-ablation.llama-3.1-8b-funding-extraction-sft-ablations
LLaMA 3.1 8B Funding Extraction SFT Ablations
Ablation study results for LoRA SFT of Meta LLaMA 3.1 8B Instruct on structured funding metadata extraction from scholarly text.
The model extracts four fields: funder_name, award_ids, funding_scheme, and award_title.
Key findings
Factor
Best config
Avg F1
Overall best
synthetic, twostage (2+1 epochs), LoRA r=64, lr=3e-5
0.588
Data type
Synthetic >> non-synthetic (+0.126 avg F1)
—
LoRA rank
r=64 > r=32 > r=16
—… See the full description on the dataset page: https://huggingface.co/datasets/cometadata/llama-3.1-8b-funding-extraction-sft-ablations.ablation-eval
Lean Proof-Ablation Eval
Syntactic proof-ablation challenges from 57 real Lean 4 repositories
(compilers, cryptography, distributed protocols, zk circuits, program logics —
see the repo list below). Each record is a (challenge, solution) pair: the
challenge is a real source file with one or more lemmas deleted and their
in-file users holed (sorry); the solution is the original file. A solver
must re-derive the deleted lemma(s) and close the holes so the file compiles.
Every… See the full description on the dataset page: https://huggingface.co/datasets/for-all-dev/ablation-eval.qwen35-08b-seqlen-ablation-0919kl3m_fw_ablationwish-engine-toolcall-id-selection-v5-ablation
wish-engine-toolcall-id-selection-v5-ablation
Targeted dataset for correcting canonical tool-ID selection mistakes with synthetic hard negatives.
Splits
train: 202
validation: 30
test: 31
Source mix
v3 generalized tool-call rows selected by eval mistake IDs
ID-selection repair supervision from real benchmark failures
synthetic hard negatives with deterministic ID perturbation + roster reorder
Generated by:… See the full description on the dataset page: https://huggingface.co/datasets/sahilmob/wish-engine-toolcall-id-selection-v5-ablation.javascript_ablationpython_ablationcommoncorpus_fw_ablationosworld-v1-hcu-ablations
OSWorld V1 HCU Ablation Trajectories
Access-controlled research package with a public dataset card and manually approved file access. Seven historical configurations, 361 non-Google-Drive tasks each, one archived trial per task. External-web tasks are retained. No V2 or new rollouts. Historical Judge observations are a separate, reference-only supplement.
Request Access
Sign in to your own Hugging Face account and request access on this dataset's
page. Requests… See the full description on the dataset page: https://huggingface.co/datasets/hqeric/osworld-v1-hcu-ablations.java_ablationphp_ablationsft_ablations_redsearcher_sftgo_ablationruby_ablationtestbed_ablationsft_ablations_redsearcher_sft_sanitizedbias_ablationsft_ablations_scientific_minimax_v1sft_ablations_bc_only_v1shotpath-boundary-cot-v7-2-output-only-ablation-data-20260728
ShotPath Boundary CoT v7.2 Output-Only Ablation
This is a controlled ablation of the v7.2 recall-adjusted Instruct data.
The same 1,356 sample IDs, images, prompts, and issue labels are retained.
The assistant target contains only the final {"items": [...]} object.
All structured reasoning before the final output is removed.
Directed no-issue rows remain capped at 5% per dimension.
Every full-review row contains at least one issue.
Images are reused from the canonical ShotPath… See the full description on the dataset page: https://huggingface.co/datasets/purefall/shotpath-boundary-cot-v7-2-output-only-ablation-data-20260728.pusht_norm4_stopreq_cot_singlenext_keystep_ablation
pusht_norm4_stopreq_cot_singlenext_keystep_ablation
BAGEL VLM-Gym SFT dataset (pusht cot ABLATION).
ABLATION: PushT all-step CoT, single-next key-step variant (matches the 14608 ablation SFT); ordered
local source dir: pusht_96_norm4_visual_nomarker_allstep_thinking_cot_stopreq_singlenext_keystep_ablation_20260607_ordered
format: trajectories/<task>/{train,test}/*.jsonl.gz (ordered by __bagel_order_key); base64 images inline.
plain & cot variants are aligned 1:1 by… See the full description on the dataset page: https://huggingface.co/datasets/novastar114/pusht_norm4_stopreq_cot_singlenext_keystep_ablation.shotpilot-matched-e-ablation-qwen3vl8b-lora
ShotPilot matched Evidence ablation
This public reproducibility archive contains three LoRA checkpoints from a strict
Evidence ablation of the ShotPilot EPA experiment. The 2,483 training rows use the
same images, prompts, ordering, hidden dimensions, Problems, and Actions as the EPA
training set. Only the assistant-side Evidence: block was removed.
Training
Base model: Qwen3-VL-8B-Instruct
Hardware: 2 x RTX 4090D, DDP
LoRA: rank 64, alpha 32, dropout 0.05… See the full description on the dataset page: https://huggingface.co/datasets/purefall/shotpilot-matched-e-ablation-qwen3vl8b-lora.moe1-naive-k-ablation-summaries