datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
atlas-16-verifier-permission-prompt-ablation
ATLAS report 16: does the orchestrator's "cannot solve" clause suppress candidate verification?
Complete raw products of the ATLAS rl-training report 16 experiment
(GitHub issue #36). Two system-prompt arms of the same model over the
same 78 fixed states, greedy decoding, one shared vLLM server.
What the experiment did
The ATLAS orchestrator's frozen system prompt contains the clause
You cannot solve the problem yourself; you decide when to explore
further and… See the full description on the dataset page: https://huggingface.co/datasets/t2ance/atlas-16-verifier-permission-prompt-ablation.T2T-Centromere-Regulatory
T2T Centromere Regulatory
Curated and released by Basepair | Follow updates on X: @BasepairSci.
Dataset Summary
The T2T Centromere Regulatory is the first comprehensive, base-pair resolution mapping of cryptic transcriptional switches and secondary structural elements across the newly sequenced Telomere-to-Telomere (T2T-CHM13 v2.0 / hs1) human centromeres.
For decades, centromeric alpha-satellite DNA (~100–200 Mb across human chromosomes) was considered… See the full description on the dataset page: https://huggingface.co/datasets/Basepair/T2T-Centromere-Regulatory.testFlexiSLM-Data-5M-t2t
FlexiSLM-Data — Text-to-Text Part (5M)
Paper: https://arxiv.org/abs/2606.31247
Demo page: https://flexislm.github.io/
Code: https://github.com/AmphionTeam/FlexiSLM
FlexiSLM-Data is a large-scale, single-turn English speech-to-speech dialogue dataset
for training FlexiSLM, a spoken language model.
This repository contains the paired prompt-and-response audio portion of the release in
WebDataset format.
Related data releases
FlexiSLM/FlexiSLM-Data-5M-t2t (this repo) provides… See the full description on the dataset page: https://huggingface.co/datasets/FlexiSLM/FlexiSLM-Data-5M-t2t.t2x
T2X (Triples-to-isiXhosa)
T2X (Triples-to-isiXhosa) is a data-to-text dataset for isiXhosa. It was constructed by translating a subset of the English WebNLG dataset into isiXhosa: each example maps a (subject, relation, object) triple from DBPedia to one or more isiXhosa sentences describing it. It can be used to train and evaluate sequence-to-sequence models for generating isiXhosa text from structured data.
Example
Triple
(South Africa, leaderName, Cyril… See the full description on the dataset page: https://huggingface.co/datasets/uctnlp/t2x.SERA-KimiK3-Django-SWEAgent-Raw-T2
SERA Kimi-K3 Django SWE-Agent - RAW T2 (second rollout)
160 raw, pre-postprocess second-rollout agent trajectories generated with the
SERA SVG pipeline (paper),
using Kimi K3 as the teacher.
Each row is an independent attempt at the synthetic PR derived from a first rollout; target_patch
holds that first-rollout patch so soft verification can be recomputed at any r.
Companion: SERA-KimiK3-Django-SWEAgent-Raw-T1.
Schema
Mirrors allenai/Sera-*-T1/T2:
column… See the full description on the dataset page: https://huggingface.co/datasets/thientrangngv/SERA-KimiK3-Django-SWEAgent-Raw-T2.SERA-GLM5.2-Django-SWEAgent-Raw-T2
SERA GLM-5.2 Django SWE-Agent - RAW T2 (second rollout, thinking enabled)
204 raw, pre-postprocess second-rollout trajectories with native GLM-5.2 reasoning traces, from the SERA SVG pipeline (paper).
Each row is an independent attempt at the synthetic PR generated from a first rollout; target_patch holds that first-rollout patch so you can recompute soft verification at any r. Companion: SERA-GLM5.2-Django-SWEAgent-Raw-T1.
Schema
Mirrors allenai/Sera-*-T1/T2:… See the full description on the dataset page: https://huggingface.co/datasets/thientrangngv/SERA-GLM5.2-Django-SWEAgent-Raw-T2.SERA-KimiK3-Django-SWEAgent-Cliff32k-T2
SERA Kimi-K3 Django SWE-Agent — Cliff-chunked T2 (second rollout)
227 training records built from 137 Kimi-K3 SWE-agent trajectories on
Django, split to fit a 32,768-token context with
CliffCompaction instead of being truncated.
Why chunked
A 100+ step agent rollout does not fit a 32k training window — 27% of the
source T2 trajectories exceed it. Truncating them throws away most of the
supervision, and trains the model on a context format it never sees at… See the full description on the dataset page: https://huggingface.co/datasets/thientrangngv/SERA-KimiK3-Django-SWEAgent-Cliff32k-T2.SERA-GLM5.2-Django-SWEAgent-T2
SERA GLM-5.2 Django SWE-Agent — T2 (second-rollout trajectories, r=0 / unverified)
165 software-engineering agent trajectories generated with the SERA SVG pipeline (paper).
Teacher: GLM-5.2 (temperature 0.6), reasoning traces preserved in <think> blocks
Harness: SWE-agent (str_replace_editor, bash, submit tools), 75-step cap, SWE-Bench Django container (django__django-7530, base commit f8fab6f9)
Stage: rollout two — a fresh agent solves the synthetic PR generated from rollout… See the full description on the dataset page: https://huggingface.co/datasets/thientrangngv/SERA-GLM5.2-Django-SWEAgent-T2.Sera-4.6-Lite-T2-v4-1000
laion/Sera-4.6-Lite-T2-v4-1000
Row-subset of allenai/Sera-4.6-Lite-T2
(the dataset upstream SERA-8B was trained on), with OpenAI tool_calls pre-rendered
into the content string as Hermes/Qwen3-style <tool_call>...</tool_call> wire tokens
and tool responses wrapped as <tool_response>...</tool_response>.
This mirrors Ai2's sera/datagen/data/postprocess/utils.py::transform_traj_hermes
(default tool_call_format: "hermes") which is the missing step between the public
Sera-4.6-Lite-T2… See the full description on the dataset page: https://huggingface.co/datasets/laion/Sera-4.6-Lite-T2-v4-1000.Sera-4.6-Lite-T2-v4-316
laion/Sera-4.6-Lite-T2-v4-316
Row-subset of allenai/Sera-4.6-Lite-T2
(the dataset upstream SERA-8B was trained on), with OpenAI tool_calls pre-rendered
into the content string as Hermes/Qwen3-style <tool_call>...</tool_call> wire tokens
and tool responses wrapped as <tool_response>...</tool_response>.
This mirrors Ai2's sera/datagen/data/postprocess/utils.py::transform_traj_hermes
(default tool_call_format: "hermes") which is the missing step between the public
Sera-4.6-Lite-T2… See the full description on the dataset page: https://huggingface.co/datasets/laion/Sera-4.6-Lite-T2-v4-316.ea-cot-t2i
EA-CoT-T2I
EA-CoT-T2I is the text-to-image (T2I) companion release of the Evaluation Agent chain-of-thought supervision data. It is published separately from the T2V-only EA-CoT-10K dataset.
The dataset contains history-conditioned next-step records distilled from multi-round T2I model-evaluation trajectories. It is designed for supervised fine-tuning of an evaluation planner that selects an evaluation tool, interprets the resulting observation, and eventually produces a… See the full description on the dataset page: https://huggingface.co/datasets/open-ea/ea-cot-t2i.llama3_t2p_reasoning_chatMMDecodingTrust-T2I-Privacysera-4.5-django-t2-recall05-toolcalls
SERA-4.5A Django T2 (Recall=0.5) Toolcalls
This dataset contains normalized multi-turn tool-calling trajectories derived from:
Source dataset: allenai/Sera-4.5A-Django-T2
Filter: line_level_recall == 0.5
Splits
train.jsonl: 6200 records
val.jsonl: 331 records
Format
Each line is a JSON object with:
id: trajectory id
messages: normalized chat/tool-call messages
metadata: includes instance_id, func_name, func_path, line_level_recall
Processing… See the full description on the dataset page: https://huggingface.co/datasets/endsky/sera-4.5-django-t2-recall05-toolcalls.