datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
continue_vs_terminate_neg_Qwen3-1.7B_DAPO-Math-en_BATCHcontinue_vs_terminate_Qwen3-1.7B_DAPO-Math-en_BATCHcontinue_vs_terminate_DAPO-Math-enDeepSeek-1.5B_dapo2k_16384_train32test0qwen3-8b-dapo-small-standard-rollouts-n8DeepSeek-1.5B_dapo14k_train1test0_part3DeepSeek-1.5B_dapo1k_16384_train32test0DeepSeek-1.5B_dapo14k_train0test64dapo-math-17k-difficulty-qwen3-1.7b-base-k16
DAPO-Math-17k difficulty under Qwen3-1.7B-Base (K=16)
For each of the 17,398 problems in the DAPO-Math-17k train set, how many of
K=16 samples from the untrained base model are correct.
The headline: 57.27% of problems are solved 0 out of 16 times, and not one
problem is solved 16 out of 16. Difficulty here is entirely one-sided.
Why count per problem instead of reporting mean accuracy
In group-relative RL (GRPO and its relatives), a prompt group whose K responses… See the full description on the dataset page: https://huggingface.co/datasets/RyanYr/dapo-math-17k-difficulty-qwen3-1.7b-base-k16.DeepSeek-1.5B_dapo14k_train1test0_part2r1-qwen7b-dapo-small-standard-rollouts-n8dapo-math-17k-qwen3-1.7b-base-n8
DAPO-Math-17k sampled with Qwen3-1.7B-Base, n=8
17398 problems from the RL training set, each sampled 8 times and scored with
the reward function the RL runs themselves used.
The point of this dataset is to be comparable with what the RL runs actually saw, so
every sampling knob is taken from the live training config or from the default that
config falls through to. Two of them are not in the config file at all and would be
wrong if guessed: top_k = -1 and min_tokens = 1.… See the full description on the dataset page: https://huggingface.co/datasets/RyanYr/dapo-math-17k-qwen3-1.7b-base-n8.dapo-filtered
dapo-filtered
A small, difficulty-banded slice of competition math, cut so that a specific policy
solves each problem rarely but not never. Built for rejection fine-tuning (RFT) and
RL experiments, where a set the model never solves gives nothing to bootstrap from and a
set it always solves gives nothing to learn.
Source: open-r1/DAPO-Math-17k-Processed
(config en). Answers are bare integers; the reward is exact-match on a \boxed{} answer.
Files
file
rows… See the full description on the dataset page: https://huggingface.co/datasets/kushasareen/dapo-filtered.continue_vs_terminate_Qwen3-1.7B_DAPO-Math-en_0-2000qwen3-14b-dapo-small-early-stop-rollouts-n8DAPO-Math-17k-easy6k-Qwen3-4B-k8DAPO-Math-17k-mixed-Qwen3-4B-k8combined_gsm8k_math_dataset_dapo_math_17k_Qwen3-4B_ntokens2048_sftdapo-3br1-qwen14b-dapo-small-standard-rollouts-n8gnosis-qwen3-8b-triviaqa-dapo-sft-2comp
Gnosis SFT — Qwen3-8B (TriviaQA + DAPO-Math)
This dataset contains SFT-style samples used to train Gnosis, a lightweight self-awareness head for LLM correctness detection.
It mixes two sources:
Math: open-r1/DAPO-Math-17k-Processed (2 model completions per question)
Trivia: mandarjoshi/trivia_qa (1 completion per question)
Each row includes the generated completion plus fields needed for correctness supervision (e.g., question/prompt, completion, and binary correctness labels).
gnosis-qwen3-4b-thinking-triviaqa-dapo-sft-2comp
Gnosis SFT — Qwen3-4B-Thinking-2507 (TriviaQA + DAPO-Math)
This dataset contains SFT-style samples used to train Gnosis, a lightweight self-awareness head for LLM correctness detection.
It mixes two sources:
Math: open-r1/DAPO-Math-17k-Processed (2 model completions per question)
Trivia: mandarjoshi/trivia_qa (1 completion per question)
Each row includes the generated completion plus fields needed for correctness supervision (e.g., question/prompt, completion, and binary… See the full description on the dataset page: https://huggingface.co/datasets/AmirhoseinGH/gnosis-qwen3-4b-thinking-triviaqa-dapo-sft-2comp.gnosis-qwen3-1_7b-triviaqa-dapo_train
Gnosis SFT — Qwen3-1.7B (TriviaQA + DAPO-Math)
Paper | Code
This dataset contains Qwen3-1.7B generated completions with binary correctness labels for training Gnosis, a lightweight self-awareness mechanism. Gnosis enables frozen LLMs to perform intrinsic self-verification by decoding signals from hidden states and attention patterns to predict the correctness of generated outputs.
It mixes two source benchmarks:
open-r1/DAPO-Math-17k-Processed (math reasoning)
mandarjoshi/trivia_qa… See the full description on the dataset page: https://huggingface.co/datasets/AmirhoseinGH/gnosis-qwen3-1_7b-triviaqa-dapo_train.autoteacher-dapo-claude-solved-coded
autoteacher-dapo-claude-solved-coded
zjhhhh/autoteacher-dapo-claude-solved (512 DAPO math problems with Claude-written, student-verified hints) augmented with a strategy coding of every hint against a compact codebook of 74 reusable problem-solving strategies.
Companion codebook dataset: zjhhhh/autoteacher-dapo-codebook.
Added columns
column
type
description
hint_code_indices
list[int]
code_ids (1–74) of the codebook strategies that form the core of the… See the full description on the dataset page: https://huggingface.co/datasets/zjhhhh/autoteacher-dapo-claude-solved-coded.Q3-1.7B-Dapo-Concqwen3-14b-dapo-small-standard-rollouts-n8Q3-1.7B-Dapo-conc-pairgnosis-gpt-oss-20b-triviaqa-dapo-sft-2comp-lowthinking
Gnosis SFT — GPT-OSS-20B (TriviaQA + DAPO-Math)
This dataset contains SFT-style samples used to train Gnosis, a lightweight self-awareness head for LLM correctness detection.
Paper: Can LLMs Predict Their Own Failures? Self-Awareness via Internal Circuits
GitHub Repository: Amirhosein-gh98/Gnosis
Dataset Description
The dataset is designed to provide supervision for detecting model failures by inspecting internal states. It mixes two primary sources:
Math: Derived… See the full description on the dataset page: https://huggingface.co/datasets/AmirhoseinGH/gnosis-gpt-oss-20b-triviaqa-dapo-sft-2comp-lowthinking.DeepSeek-1.5B_dapo2k_16384_train32test0_justifiedr1-qwen14b-dapo-small-early-stop-rollouts-n8