datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Dolci-Instruct-RL-Completions
Dolci Instruct RL Completions
Instruction-following completions sampled from OLMo-3-7B-Instruct with teacher logits for knowledge distillation.
Description
This dataset contains instruction-completion pairs with pre-computed teacher logits from OLMo-3-7B-Instruct. Designed for training smaller student models via KL-divergence distillation.
Generation
Completions were sampled from allenai/OLMo-3-7B-Instruct on instruction prompts. For each token position, we… See the full description on the dataset page: https://huggingface.co/datasets/hbfreed/Dolci-Instruct-RL-Completions.jb-completions
JB-Completions Dataset: Base Model Safety Evals
Overview
JB-Completions is a dataset designed for evaluating the harmfulness of base language models (i.e., completion/non-instruction-fine-tuned LLMs). This dataset contains pairs of harmful prompts and their corresponding completions, allowing researchers to assess how base models respond to potentially harmful inputs. See our paper on Safety Pretraining for more details!
Dataset Structure
The dataset… See the full description on the dataset page: https://huggingface.co/datasets/locuslab/jb-completions.wildchat-glm53-format-completions
WildChat format completions
9,975 GLM-5.3 answers across 29 parseable formats. Each answer passed its
contract verifier. Failed answers were resampled with the same prompt until one
passed; no semantic judge or answer repair was used.
This is the final release from a 10,000-prompt run; 25 unfinished prompts were excluded. It stores
answers, format instructions, exact contracts, and pinned
WildChat-4.8M references—not
the source prompts or conversations. All rows are in the train… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/wildchat-glm53-format-completions.Qwen3.8-27B-thinking-completions
Qwen3.8-27B thinking-mode completions
17,022 prompts from public chat, math and code datasets, each answered once by Qwen3.8-27B (FP8 checkpoint) in thinking mode with its recommended sampling settings (54M completion tokens). Every sample has the reasoning trace and the final answer, as text and as the exact token ids.
The set was generated to train speculative-decoding drafters for this model, so it records the model's own sampled distribution rather than greedy output or… See the full description on the dataset page: https://huggingface.co/datasets/JonasLoos/Qwen3.8-27B-thinking-completions.tiny-vintage-completions
Tiny vintage completions
Synthetic vintage texts, with a cutoff date for year 1900.
Based on unique 2-3 word seeds, extracted from croqaz/Vintage-v1, croqaz/Vintage-v2 and Haykgrigorian/English-historical-corpus-1800-1875.
Check the files seeds1.txt and seeds2.txt.
Generated by TypeWriter-7B-base and Talkie-13B-base completions.
Citation
If you find this dataset valuable, please consider citing:
@misc{Tiny-vintage-completions,
title = {Tiny vintage completions}… See the full description on the dataset page: https://huggingface.co/datasets/croqaz/tiny-vintage-completions.heretic-completions
Heretic Completions
Model completions used as SFT targets for a refusal-abliteration LoRA study.
Each row pairs a prompt from a red-teaming / over-refusal benchmark with a
completion from a refusal-removed ("heretic" / abliterated) model.
Safety notice. This is a private research dataset. Many completions
comply with harmful or dual-use requests by design, so the refusal signal
can be measured and abliteration studied. Do not redistribute or use outside
authorized safety… See the full description on the dataset page: https://huggingface.co/datasets/noahrossi/heretic-completions.gemma-4-e2b-deception-behavior-completions
Gemma-4-E2B deception & behavior completions
Consolidated 910-row corpus of (scenario prompt + Gemma-4-E2B-generated completion) pairs from earlier mechanistic-interpretability experiments. Each row captures the prompt the model saw and the text it actually produced; for a subset, Claude-Haiku-4-5 judge verdicts and SAE-feature labels are included.
The corpus is meant to be used as activation-extraction input for downstream interpretability work — Natural Language Autoencoder (NLA)… See the full description on the dataset page: https://huggingface.co/datasets/Solshine/gemma-4-e2b-deception-behavior-completions.rr-circuit-breakers-attack-completions
RR (Circuit Breakers) attack completions with three-judge scores
This dataset bundles attack completions generated against
GraySwanAI/Llama-3-8B-Instruct-RR
(the "circuit breakers" defense), each scored by three independent judges:
local:strongreject (Lin et al., StrongREJECT classifier — most permissive)
local:harmbench (HarmBench classifier — middle)
local:gpt_oss (gpt-oss-safeguard-20b — strictest)
Headline finding: judges DISAGREE dramatically on… See the full description on the dataset page: https://huggingface.co/datasets/samuelsimko/rr-circuit-breakers-attack-completions.