datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
Qwen3.5-4B-Base
juiceb0xc0de/Qwen3.5-4B-Base
A brain atlas for Qwen/Qwen3.5-4B-Base, a 32-layer hybrid that runs linear attention on 24 layers and full attention on the other 8. This is not a chat dataset or a benchmark. It is an internal-mechanics map built by running activations through a corpus of prompts and scoring what each layer, component, head, and feature direction is doing.
This is a base model, before any instruction tuning, so whatever structure shows up here was put there by… See the full description on the dataset page: https://huggingface.co/datasets/juiceb0xc0de/Qwen3.5-4B-Base.qwen35-4b
qwen35-4b
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.38203125
Action score: 0.4375
Valid samples: 320/320
appworld-qwen35-4b-9b-s_signal_6-epoch4-iter1
appworld-qwen35-4b-9b-s_signal_6-epoch4-iter1
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.3953125
Action score: 0.446875
Valid samples: 320/320
qwen35-4b-reeval3
qwen35-4b-reeval3
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.3859375
Action score: 0.4125
Valid samples: 320/320
qwen35-4b-reeval2
qwen35-4b-reeval2
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.384375
Action score: 0.4265625
Valid samples: 320/320
qwen35-4b-reeval1
qwen35-4b-reeval1
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.39296875
Action score: 0.4203125
Valid samples: 319/320
appworld-qwen35-4b-agent-rl-epoch3
appworld-qwen35-4b-agent-rl-epoch3
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.45859375
Action score: 0.475
Valid samples: 320/320
qwen35-4b-reeval4
qwen35-4b-reeval4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.3984375
Action score: 0.4265625
Valid samples: 319/320
appworld-qwen35-4b-manysource-2k-newprompt-solvability-junhee-epoch8-tmp01-reeval1
appworld-qwen35-4b-manysource-2k-newprompt-solvability-junhee-epoch8-tmp01-reeval1
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.40546875
Action score: 0.475
Valid samples: 320/320
appworld-qwen35-4b-manysource-2k-newprompt-solvability-junhee-epoch8-reeval1
appworld-qwen35-4b-manysource-2k-newprompt-solvability-junhee-epoch8-reeval1
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.4046875
Action score: 0.4703125
Valid samples: 320/320
appworld-qwen35-4b-agent-rl-epoch3-reeval1
appworld-qwen35-4b-agent-rl-epoch3-reeval1
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.4578125
Action score: 0.4921875
Valid samples: 320/320
appworld-qwen35-4b-manysource-2k-newprompt-solvability-junhee-epoch8
appworld-qwen35-4b-manysource-2k-newprompt-solvability-junhee-epoch8
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.39921875
Action score: 0.44375
Valid samples: 320/320
appworld-qwen35-4b-manysource-2k-newprompt-solvability-junhee-epoch8-t01
appworld-qwen35-4b-manysource-2k-newprompt-solvability-junhee-epoch8-t01
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.38359375
Action score: 0.4703125
Valid samples: 320/320
appworld-qwen35-4b-9b-s_signal_5-epoch4-iter1-reeval1
appworld-qwen35-4b-9b-s_signal_5-epoch4-iter1-reeval1
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.41328125
Action score: 0.4359375
Valid samples: 320/320
appworld-qwen35-4b-9b-s_signal_5-epoch4-iter1
appworld-qwen35-4b-9b-s_signal_5-epoch4-iter1
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.41953125
Action score: 0.4515625
Valid samples: 320/320
appworld-qwen35-4b-total-237-audited-jh-epoch2
appworld-qwen35-4b-total-237-audited-jh-epoch2
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.3640625
Action score: 0.4328125
Valid samples: 320/320
appworld-qwen35-4b-total-237-audited-jh-epoch6
appworld-qwen35-4b-total-237-audited-jh-epoch6
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.37578125
Action score: 0.421875
Valid samples: 320/320
appworld-qwen35-4b-total-237-audited-jh-epoch8
appworld-qwen35-4b-total-237-audited-jh-epoch8
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.384375
Action score: 0.4390625
Valid samples: 320/320
qwen35-4b-filter-s_signal5-200-qwen38-27b-newprompt-4k-epoch4
qwen35-4b-filter-s_signal5-200-qwen38-27b-newprompt-4k-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.3890625
Action score: 0.4359375
Valid samples: 320/320
qwen35-4b-filter-solvability-200-qwen38-27b-newprompt-4k-epoch4
qwen35-4b-filter-solvability-200-qwen38-27b-newprompt-4k-epoch4
Portable process-evaluation output. metadata.json is the lightweight source
for aggregate results; the JSONL files are directly loadable; and
artifacts.tar.gz losslessly preserves the original run directory.
Reasoning score: 0.40234375
Action score: 0.421875
Valid samples: 320/320
tiktok-videos-4b
TikTok Videos: 4.5 billion posts dataset
Step-by-step guide and access to the scraper code:
tiktok-api.seeksocial.io.
4.5 billion TikTok video records with captions, engagement counts, sound
identifiers and timing. Collected from TikTok's mobile API over roughly three
weeks. Every content_id appears exactly once.
This is the largest public TikTok dataset I am aware of. It is released as-is,
for research.
What is in it
27 Parquet files, zstd compressed, about 289… See the full description on the dataset page: https://huggingface.co/datasets/kwakuobeng/tiktok-videos-4b.tiktok-videos-4b
TikTok Videos: 4.5 billion posts dataset
Step-by-step guide and access to the scraper code:
tiktok-api.seeksocial.io.
4.5 billion TikTok video records with captions, engagement counts, sound
identifiers and timing. Collected from TikTok's mobile API over roughly three
weeks. Every content_id appears exactly once.
This is the largest public TikTok dataset I am aware of. It is released as-is,
for research.
What is in it
27 Parquet files, zstd compressed, about 289… See the full description on the dataset page: https://huggingface.co/datasets/blaccastro/tiktok-videos-4b.essay-vocab-range-qwen3.5-4b-trl-completions
TRL Completion logs
This dataset contains the completions generated during training using trl.
Find the trained model at https://huggingface.co/bihungba1101/essay-vocab-range-qwen3.5-4b-grpo.
The completions are stored in parquet files, and each file contains the completions for a single step of training (depending on the logging_steps argument).
Each file contains the following columns:
step: the step of training
prompt: the prompt used to generate the completion
completion: the… See the full description on the dataset page: https://huggingface.co/datasets/bihungba1101/essay-vocab-range-qwen3.5-4b-trl-completions.tiktok-videos-4b
TikTok Videos: 4.5 billion posts with engagement metrics
4.5 billion TikTok video records with captions, engagement counts, sound
identifiers and timing. Collected from TikTok's mobile API over roughly three
weeks. Every content_id appears exactly once.
This is the largest public TikTok dataset I am aware of. It is released as-is,
for research.
What is in it
27 Parquet files, zstd compressed, about 289 GB in total. One row per video.
Column
Type
Description… See the full description on the dataset page: https://huggingface.co/datasets/dams2005/tiktok-videos-4b.grammar-accuracy-qwen3.5-4b-trl-completions
TRL Completion logs
This dataset contains the completions generated during training using trl.
Find the trained model at https://huggingface.co/bihungba1101/grammar-accuracy-qwen3.5-4b-grpo.
The completions are stored in parquet files, and each file contains the completions for a single step of training (depending on the logging_steps argument).
Each file contains the following columns:
step: the step of training
prompt: the prompt used to generate the completion
completion:… See the full description on the dataset page: https://huggingface.co/datasets/bihungba1101/grammar-accuracy-qwen3.5-4b-trl-completions.tiktok-videos-4b
Mirror of kuben-developer/tiktok-videos-4b, snapshot 2026-09-08. All credit to the original author; same research-use license applies.
TikTok Videos: 4.5 billion posts with engagement metrics
4.5 billion TikTok video records with captions, engagement counts, sound
identifiers and timing. Collected from TikTok's mobile API over roughly three
weeks. Every content_id appears exactly once.
This is the largest public TikTok dataset I am aware of. It is released as-is,
for research.… See the full description on the dataset page: https://huggingface.co/datasets/merway/tiktok-videos-4b.maxrl_qwen3_4B_base_polaris_rollouts
MaxRL Qwen3-4B-Base training rollouts (POLARIS math prompts)
Every training rollout from an online RL run, with exact token ids, sampling
log-probs, and raw rewards — usable as a replay buffer to study off-policy RL
for LLM reasoning completely offline.
The run: Qwen3-4B-Base trained with the maxRL advantage estimator
(A = (r - mean)/(mean + eps), group mean over 16 rollouts per prompt;
maxRL paper) and a pure REINFORCE loss
(L = -A * log pi; no importance ratio, no clipping, no… See the full description on the dataset page: https://huggingface.co/datasets/ftajwar/maxrl_qwen3_4B_base_polaris_rollouts.essay-grammar-range-qwen3.5-4b-trl-completions
TRL Completion logs
This dataset contains the completions generated during training using trl.
Find the trained model at https://huggingface.co/bihungba1101/essay-grammar-range-qwen3.5-4b-grpo.
The completions are stored in parquet files, and each file contains the completions for a single step of training (depending on the logging_steps argument).
Each file contains the following columns:
step: the step of training
prompt: the prompt used to generate the completion
completion: the… See the full description on the dataset page: https://huggingface.co/datasets/bihungba1101/essay-grammar-range-qwen3.5-4b-trl-completions.tiktok-videos-4b
TikTok Videos: 4.5 billion posts dataset
Step-by-step guide and access to the scraper code:
tiktok-api.seeksocial.io.
4.5 billion TikTok video records with captions, engagement counts, sound
identifiers and timing. Collected from TikTok's mobile API over roughly three
weeks. Every content_id appears exactly once.
This is the largest public TikTok dataset I am aware of. It is released as-is,
for research.
What is in it
27 Parquet files, zstd compressed, about 289… See the full description on the dataset page: https://huggingface.co/datasets/KOM-00/tiktok-videos-4b.4B-Instruct-reason-only.stride-1.k-8.statml-arxiv.qwen3-ids
4B-Instruct-reason-only.stride-1.k-8.statml-arxiv.qwen3-ids
Thoughts for next-token prediction on k=8 token chunks of JackHsieh/statML-arxiv-40M-20M, generated by
Qwen3-4B-Instruct-2507. Each thought is a few dense sentences of reasoning about the next
8 tokens after a cut, written from the document prefix alone — the generator never sees the
continuation. Stored thought_text includes the <thought>/</thought> wrapper.
This is the small-generator parity counterpart of… See the full description on the dataset page: https://huggingface.co/datasets/JackHsieh/4B-Instruct-reason-only.stride-1.k-8.statml-arxiv.qwen3-ids.