datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
snowball-replay-index
Snowball replay index
This dataset is a compact membership and ordering index for an approximate replay of Snowball's 10,372,343,704,053-token
data store. It contains no source text or token arrays. The 6,301 Parquet files contain three columns:
source_id: logical source key; join it to the source_id field in sources.json
document_id: the retained XXH3-128 content hash as 16 bytes
bucket_id: domain_cluster * 5 + quality_bucket
Document join contract
document_id… See the full description on the dataset page: https://huggingface.co/datasets/open-athena/snowball-replay-index.climbmix-seed42-10b-replay
ClimbMix seed-42 training replay, approximately 10B tokens
manifest.json is the authoritative export status: only state: ready means
construction is complete. It has not been uploaded to Hugging Face.
This is a frozen training input stream, not a new raw-text mixture. It uses
the existing OptimalScale/ClimbMix revision and the exact tokenizer pinned in
the manifest. Documents are shuffled with seed 42 and split into the original
eight contiguous virtual-rank streams before… See the full description on the dataset page: https://huggingface.co/datasets/ArchloverLRZ/climbmix-seed42-10b-replay.2026-07-31-tulu3-replay-80-pct-qwen36-mixture
TULU3 replay slice — 80% of the Qwen3.6-27B difficult-advice mixture
The replay half of the 20/80 training mixture used for the Qwen3.6-27B
difficult-advice arms: 1,878 conversations, 1,194,548 tokens, exactly
80.0% of that mixture. The other 20% is difficult-advice data and is not
included here.
Published so the replay portion can be reused or audited on its own. Sampled from
allenai/tulu-3-sft-mixture
with seed=0, keeping only conversations that end on an assistant turn… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-07-31-tulu3-replay-80-pct-qwen36-mixture.deepseek-v4-flash-swebench-replay
deepseek-v4-flash-swebench-replay
中文
这是一个 DeepSeek V4 Flash 在 SWE-bench 上的 agentic replay 数据集仓库。
它的目标是让使用者不需要部署 SWE-bench，也不需要复现 Docker/benchmark 环境，就可以直接查看和重放模型的多轮推理与工具调用轨迹。
当前包含的数据
verified_agentic
lite_agentic
当前不包含的数据
单轮 single-turn trace
verified_mini_agentic（当前本地仅完成 31/50，因此不纳入首版）
分数汇总
verified_agentic: 354 / 500, Acc/Pass@1 = 70.8
lite_agentic: 182 / 300, Acc/Pass@1 = 60.67
数据来源
这些轨迹由… See the full description on the dataset page: https://huggingface.co/datasets/fxiao0369/deepseek-v4-flash-swebench-replay.replay-gap-trajectories
The Replay Gap: Branched Agent Trajectories
Counterfactual ("branched") agent rollouts for studying per-step model
switching in LLM agents, from the paper The Replay Gap: Static Evaluation of
Model Switching in LLM Agents Scores the Wrong World (Efficient Reasoning
Workshop @ COLM 2026).
Routing benchmarks score routers by replaying logged model outputs. In a
multi-step agent that is unsound: swap the model at step k and the rest of
the trajectory diverges. This dataset contains… See the full description on the dataset page: https://huggingface.co/datasets/ashritha0907/replay-gap-trajectories.ReplayDPO
ReplayDPO
Preference pairs for DPO/DMPO training of tool-using agents (Qwen3.5-9B baseline). Pairs were extracted from a per-block judge + multi-branch replay pipeline run on StableToolBench (G1, G2) trees and HotpotQA failure replays.
Files
file
pairs
filter
dpo_pairs_m05.jsonl (v1)
573
min_margin=0.05 + cache-miss soft filter
dpo_pairs_m10.jsonl (v1)
391
min_margin=0.10 + cache-miss soft filter
dpo_pairs_m05_v2.jsonl
329
v1 + drop HotpotQA + drop… See the full description on the dataset page: https://huggingface.co/datasets/Naclin/ReplayDPO.dclm-replay.seq-4096.n-262144-olmo3
dclm-replay.seq-4096.n-262144-olmo3
OLMo 3-tokenized CPT-replay sequences for prestar, the OLMo analogue of JackHsieh/dclm-replay.seq-4096.tokens-32B (Qwen3).
Source: mlfoundations/dclm-baseline-1.0, pin global-shard_01_of_10/local-shard_0_of_10/*.jsonl.zst.
Tokenizer: allenai/Olmo-3-1025-7B; EOD token id 100257 (<|endoftext|>).
262,144 sequences of exactly 4096 tokens each (docs concatenated and packed; EOD-separated).
Same builder/pin as the Qwen3 replay — corpus is the same… See the full description on the dataset page: https://huggingface.co/datasets/JackHsieh/dclm-replay.seq-4096.n-262144-olmo3.deepseek-v4-pro-swebench-replay
deepseek-v4-pro-swebench-replay
中文
这是一个 DeepSeek V4 Pro 在 SWE-bench 上的 agentic replay 数据集仓库。
用户不需要部署 SWE-bench，也不需要复现 Docker/benchmark 环境，即可直接查看和重放模型的多轮推理、工具调用和最终 patch 轨迹。
数据来源
使用 EvalScope 收集
使用 EvalScope 中的 official SWE-bench agentic benchmark
模型：deepseek-v4-pro
运行模式：thinking + toolcall agentic loop
评测环境：rootless Docker + SWE-bench 官方 benchmark containers
数据范围
verified_agentic: 500 / 500 traces
lite_agentic: 300 / 300 traces… See the full description on the dataset page: https://huggingface.co/datasets/fxiao0369/deepseek-v4-pro-swebench-replay.GLM51-swebench-replay
GLM51-swebench-replay
中文
这是一个 GLM-5.1 在 SWE-bench 上的 agentic replay 数据集仓库。目标是让使用者不需要部署 SWE-bench，也不需要复现 Docker/benchmark 环境，就可以直接查看和重放模型的多轮推理、工具调用和最终 patch。
数据来源
轨迹使用 EvalScope 收集，benchmark 使用 EvalScope 中的 official SWE-bench agentic 数据集：
swe_bench_verified_agentic
swe_bench_lite_agentic
运行时使用 SWE-bench 官方容器镜像，收集形态为 agentic + toolcall。模型调用使用 GLM-5.1 的 OpenAI-compatible 接口。
分数汇总
verified_agentic: 363 / 500, Acc/Pass@1 = 72.6… See the full description on the dataset page: https://huggingface.co/datasets/fxiao0369/GLM51-swebench-replay.kanitakorn-v67-replay-guard-clean
Kanitakorn v67 Replay Guard Clean SFT
Training-ready backup replay package for preserving math, code,
instruction-following, and bilingual short-reasoning behavior. This is not the
primary ThaiExam-gain dataset. Use it only if a Thai overlay improves ThaiExam
but regresses side metrics such as MATH100/AIME/LiveCodeBench smoke tests.
Contents
train.jsonl: 172 SFT message rows.
manifest.json: provenance, counts, hashes, and recommended use.
README.md: this dataset… See the full description on the dataset page: https://huggingface.co/datasets/Jnx03/kanitakorn-v67-replay-guard-clean.molmo2-replay-500k
molmo2-replay-100k — diverse anti-forgetting replay mix (text + image + short video)
A ~100k diverse subset sampled from the molmo2 academic SFT mixture, meant to be mixed into another
training run as a replay set to prevent catastrophic forgetting of general language / image / video
abilities. It deliberately spans many tasks (per-dataset caps for diversity) and excludes long
videos (only short clips; videos > 10 MB filtered out) so it stays light on context length.… See the full description on the dataset page: https://huggingface.co/datasets/weikaih/molmo2-replay-500k.browsecomp-plus-glm52-fp8-trie-event-replay
GLM-5.2 FP8 BrowseComp-Plus Trie Event Replay
This manually gated dataset contains a captured BrowseComp-Plus agent workload
served by GLM-5.2 FP8 on SGLang with CPU BM25 retrieval and evaluation
concurrency eight: all 830 benchmark queries, one closed terminal trajectory
each, recorded as Trie schema-v2 causal event streams and directly replayable
against any OpenAI-compatible inference endpoint.
Quick start: download, cd, replay
hf download… See the full description on the dataset page: https://huggingface.co/datasets/weili-0234/browsecomp-plus-glm52-fp8-trie-event-replay.