datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
GLM-5.3-Flash-calibration-activations-v1
GLM-5.3-Flash calibration activations v1 (BF16, natural routing)
Per-layer block-input activations of zai-org/GLM-5.3-Flash-BF16 @ b1967181 over 92x2048
tokens of the exllamav3 standard_cal_data corpus (pinned): per context, layer_NNN.attn_in
and layer_NNN.mlp_in (bf16, post-norm linear inputs; mlp_in is the router + expert gate/up
input) and layer_NNN.router_logits (fp32, natural top-8 routing ground truth).
Per-expert Hessians E[xx^T], routing statistics and down-proj inputs… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/GLM-5.3-Flash-calibration-activations-v1.glm53-flash-fidelity-root-v1
fidelity--glm53flash.malaiwah.root.bf16
A root fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from zai-org/GLM-5.3-Flash-BF16.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits after it). Same cut… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm53-flash-fidelity-root-v1.glm53-flash-fidelity-exl3-tr3-6bpw-v1
fidelity--glm53flash.malaiwah.quant.tr3-6bpw
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from malaiwah/GLM-5.3-Flash-TR3-6bpw.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits after… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm53-flash-fidelity-exl3-tr3-6bpw-v1.DeepSeek-V4-Flash-0731-Teacher-Distillation-40513x
DeepSeek V4 Flash 0731 Teacher Distillation — 40,513 Retained Rows
Teacher-distillation corpus generated with
deepseek-ai/DeepSeek-V4-Flash-0731.
The original manifest contained 45,000 unique seeds.
Following generation, QC, retry-based repair, quarantine auditing,
and recovery adjudication, 40,513 rows were retained.
Composition
Bucket
Rows
Coding
5,601
Agentic
9,982
Cyber blue
13,000
Controlled cyber red
6,999
Tool use
4,931
Total
40,513… See the full description on the dataset page: https://huggingface.co/datasets/trjxter/DeepSeek-V4-Flash-0731-Teacher-Distillation-40513x.glm53-flash-fidelity-fp8-v1
fidelity--glm53flash.malaiwah.quant.official-fp8
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from zai-org/GLM-5.3-Flash.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already sits after it).… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm53-flash-fidelity-fp8-v1.swebench-verified-deepseek-v4-flash-failure-analysis
SWE-bench Verified runs & failure analysis — DeepSeek-V4-flash (local) × mini-swe-agent
Per-instance analysis of SWE-bench Verified runs of a locally-served DeepSeek-V4-flash model
driven by mini-swe-agent, graded with the official
SWE-bench harness. Each instance carries the full agent trajectory, a readable transcript, the
submitted patch, the harness test output, deterministic metrics, and a hand-verified qualitative
root-cause diagnosis.
Current numbers (resolve rates… See the full description on the dataset page: https://huggingface.co/datasets/daaain/swebench-verified-deepseek-v4-flash-failure-analysis.ESNLIR-dataset
ESNLIR
A Spanish multi-genre NLI dataset with explicit causal annotation. 4.4M training pairs drawn
from 34 corpora across 8 genres, with a fourth label — reasoning — for cause–effect relationships
that the usual entailment / contradiction / neutral scheme cannot express.
ESNLIR: Expanding Spanish NLI Benchmarks with Multi-Genre and Causal Annotation
Johan R. Portela, Nicolás Pérez-Terán, Rubén Manrique — Universidad de los Andes, Bogotá
Applied Informatics, Springer, 2026, pp.… See the full description on the dataset page: https://huggingface.co/datasets/Flaglab/ESNLIR-dataset.agent-traces-flashmaniaswival.dev agent traces for single-shot "Reimplement trackmania" prompt using Gemini-3.5-flash.
Generated code: https://github.com/dip-proto/flashmania
k3-sft-cc0-flan
Dataset Card for K3 SFT CC0 FLAN
844-row Kimi K3 synthetic instruction-tuning shard built from DPI-traced CC0/public-domain
FLAN prompts in the Tülu mix. Four overlapping Hub configs expose different cohort
views; adaptive is the recommended default for quality-conscious SFT mixing.
Dataset Details
Curated by: Training Datasmith
Teacher: kimi-k3 via deltafin (local inference)
Languages: English prompts; translation pairs include German, Spanish, Czech, Igbo… See the full description on the dataset page: https://huggingface.co/datasets/Training-Datasmith/k3-sft-cc0-flan.RAGPulse
RAGPulse: A Real-World RAG Workload Trace to Optimize RAG Serving Systems
🌐 Github Link |
🤗 Workload Trace |
📑 Arxiv Paper |
🤖 How to use?
RAGPulse is a real-world RAG workload trace collected from an university-wide Q&A service scenario. The system has been serving over 40,000 students and faculties since April 2024, providing intelligent policy Q&A services. The trace contains a total of 7,106 records entries, sampled from one week of our Q&A service.
Dataset Details… See the full description on the dataset page: https://huggingface.co/datasets/flashserve/RAGPulse.glm53-flash-fidelity-exl3-tr3-4bpw-miaailab-v1
fidelity--glm53flash.malaiwah.quant.miaailab-4bpw
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm53-flash-fidelity-exl3-tr3-4bpw-miaailab-v1.prompt-policy-memory-v0
Prompt Policy Memory v0
Synthetic profile-memory data: 100 training sessions from10users;20test sessions from2fresh users. Test users were generated after the GRPO checkpoint was frozen and must not be used for training or tuning.
Each row includes cumulative plain-text session input, a canonical plain-text key:value reference, chat messages, and evaluator-only target data. messages can be used for supervised fine-tuning. The reference contains all currently revealed facts; it… See the full description on the dataset page: https://huggingface.co/datasets/flavianv/prompt-policy-memory-v0.lm-eval-results-nbeerbower-Flammen-Bophades-7B-private
Dataset Card for Evaluation run of nbeerbower/Flammen-Bophades-7B
Dataset automatically created during the evaluation run of model nbeerbower/Flammen-Bophades-7B
The dataset is composed of 62 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 2 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest… See the full description on the dataset page: https://huggingface.co/datasets/nyu-dice-lab/lm-eval-results-nbeerbower-Flammen-Bophades-7B-private.llm_timeline_deepseek_v4_flash-pi
Coding agent session traces
This dataset contains coding agent session traces collected while working on LLM Timeline web app using the prompt from coding-agent-bench-prompts
glm53-flash-fidelity-exl3-wrld-k3-v1
fidelity--glm53flash.malaiwah.quant.exl3-wrld-k3
A quant fidelity dataset in hidden form, produced by engines/tools/hf_capture.py from wrldsuksgo2mars/GLM-5.3-Flash-EXL3-K3-v1.
The cut
the final hidden state handed to lm_head -- after the text model's final norm and immediately before the head matmul -- captured as the head module's input via torch.nn.Module.register_forward_pre_hook; replay applies the head ONLY (no final norm at replay time: the capture already… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/glm53-flash-fidelity-exl3-wrld-k3-v1.deepseek-v4-flash-swe-cot
DeepSeek-V4-Flash SWE Agent Trajectories (with raw chain-of-thought)
795 multi-turn software-engineering agent trajectories generated by
DeepSeek-V4-Flash-0731 at reasoning_effort=max, each one executed in a real
repository inside an isolated container and verified by running the repository's own
tests. 469 are verified-correct.
Every assistant turn preserves reasoning_content — the model's raw chain-of-thought,
not a summary. That is the point of this dataset: the DeepSeek API… See the full description on the dataset page: https://huggingface.co/datasets/blythet/deepseek-v4-flash-swe-cot.tomasmcm__sky-t1-coder-32b-flash-details
Dataset Card for Evaluation run of tomasmcm/sky-t1-coder-32b-flash
Dataset automatically created during the evaluation run of model tomasmcm/sky-t1-coder-32b-flash
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/tomasmcm__sky-t1-coder-32b-flash-details.repro-flat-minima-and-generalization-insights-from-stochastic-convex-optimization-traces
Agent traces
Agent sessions published from a Trackio Logbook.
swe-verified-gemini3-flash-trajectories
SWE-bench Verified — Gemini-3-flash agent trajectories (graded, 3 samples/instance)
Agent trajectories from gemini-3-flash-preview (high reasoning, temperature 0.8) run with the
OpenHands agent on SWE-bench Verified, in Modal sandboxes. For each of 100 instances
we sampled multiple trajectories and graded them with the SWE-bench harness; this dataset holds the
3 graded samples per instance = 296 trajectories, 198 resolved (67%).
pass@1 ≈ 66/100, pass@3 (oracle) = 73/100.… See the full description on the dataset page: https://huggingface.co/datasets/tarsur385/swe-verified-gemini3-flash-trajectories.repro-flashoptim-Wfe1iJocjF-traces
Agent traces
Agent sessions published from a Trackio Logbook.
arc_agi_1_augmentedStep-3.5-Flash-Instruct-EmMcts
Step-3.5-Flash-Instruct-EmMcts
Preference (chosen / rejected) dataset generated with an Empirical-MCTS (Em-Mcts) rollout pipeline
and scored by a reward model. Every sample contains a higher-quality chosen response and a
lower-quality rejected response for the same prompt, making it suitable for DPO / preference
optimization and reward-model training.
Overview
Records: 4,959
Format: JSON Lines (one JSON object per line)
Language: English
Generation model:… See the full description on the dataset page: https://huggingface.co/datasets/Minami-su/Step-3.5-Flash-Instruct-EmMcts.UltraSteer-v0-flatNote 0: This is an flattened version of the dataset where multi-turn samples were split into multiple dataset lines. If you would like the original unflattened version before splitting samples into multiple lines please use UltraSteer-v0.
UltraSteer: A Massive Collection of Multi-Turn Dialogue with Fine-Grained Labels
UltraSteer is a large-scale dataset of single- and multi-turn dialogue with fine-grained labels produced by Nvidia's Llama2-13B-SteerLM-RM reward model using the NeMo… See the full description on the dataset page: https://huggingface.co/datasets/Avelina/UltraSteer-v0-flat.google__flan-t5-large-details
Dataset Card for Evaluation run of google/flan-t5-large
Dataset automatically created during the evaluation run of model google/flan-t5-large
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.
An additional… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/google__flan-t5-large-details.human_fall_flat_recordings_01
人类一败涂地 raw recordings
This dataset contains raw game recordings managed by Game Data Platform. Access requests require manual approval.
Game ID: game_8b46e9573d7ed0a5597c311de5d10d53
Collection: general (泛数据)
Recordings: 22
Layout: recordings/<recording_id>/<raw component>
Grill-preprod-v1_chatML
Grill-v1
This dataset contains ChatML formatted conversation between a human and AI. The goal is simulating a preproduction version of flammen.ai in order to train the Mahou series of models.
Schema
chatID - index linking conversations
idx - index of response in a conversation
prompt - ChatML formatted conversation including flammen.ai system prompt
chosen - accepted response
rejected - rejected response (for DPO)
Method
SillyTavern was used as the frontend… See the full description on the dataset page: https://huggingface.co/datasets/flammenai/Grill-preprod-v1_chatML.google__flan-t5-small-details
Dataset Card for Evaluation run of google/flan-t5-small
Dataset automatically created during the evaluation run of model google/flan-t5-small
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.
An additional… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/google__flan-t5-small-details.google__flan-t5-base-details
Dataset Card for Evaluation run of google/flan-t5-base
Dataset automatically created during the evaluation run of model google/flan-t5-base
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.
An additional… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/google__flan-t5-base-details.google__flan-ul2-details
Dataset Card for Evaluation run of google/flan-ul2
Dataset automatically created during the evaluation run of model google/flan-ul2
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing to the latest results.
An additional configuration… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/google__flan-ul2-details.BlendNet
📚 BlendNet
The dataset contains $12k$ samples. To balance cost savings with data quality and scale, we manually annotated $2k$ samples and used GPT-4o to annotate the remaining $10k$ samples.
For more details, please visit our GitHub repository or refer to our arXiv paper.
📖 Citation
@misc{du2024blenderllmtraininglargelanguage,
title={BlenderLLM: Training Large Language Models for Computer-Aided Design with Self-improvement},
author={Yuhao Du and… See the full description on the dataset page: https://huggingface.co/datasets/FlandreScarlet123/BlendNet.