Skip to content

Green0-0/llm_datasets

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

36 Commits
 
 

Repository files navigation

LLM Datasets

A collection of high quality datasets from huggingface.

Pretraining Corpus:

Highly filtered pretraining datasets.

Warning: Writing is extremely underrepresented. I am in the process of looking for writing datasets.

Both finepdfs-edu and finewiki contain dense multilingual data that has been collected into language-specific subsets (en for eng finewiki, eng_Latn for eng finepdfs). Finewiki appears to be organized into markdown, finepdfs-edu does not appear particularly formatted. In finepdfs-edu, the list fw_edu_scores should contain the education scores.

Contains a generic mixture of exerts from websites. The language is under the 'language' column, score is under the 'score' column. Hopefully a fineweb2-edu will be out soon to replace this.

An alternative to fineweb-edu by openbmb (ultrachat, minicpm authors). Language subsets (en for eng) and 'score' column.

Compared to finemath which is full of noise and honestly quite terrible, l2 is quite decent for a filtered math dataset.

Surprisingly decent.

https://huggingface.co/datasets/allenai/dolma3_longmino_mix-100B-1125

Could be good for long context.

WIP, worth paying attention to:

Honorable mentions (unfiltered):

Midtraining:

These datasets can be slotted into a pretraining run at the end for curriculum learning or mixed throughout. Remember that midtraining datasets must be very large but can be lower quality; SFT is the opposite.

Warning: I extensively cite NVIDIA datasets which have been known to produce extremely dry models.

Nemotron Cascade is extremely extensive, with the one issue being that responses are generated with Deepseek R1/V3 instead of newer models such as GLM 5 or even DSV3.2 (thus I would not recommend it for SFT).

These are a bit more niche compared to Nemotron-Cascade-SFT, and might be biased towards certain domains.

Also note that in my opinion, Kimi K2.5 is way better than Kimi K2, and GLM 5 is way better than 4.6, so these datasets have depreciated quite a bit.

SFT - Conversational:

Generic multiturn conversational data for training.

Honorable Mentions:

These three are human written, but are quite outdated and low quality compared to synthetic LLM data.

SFT - Math/Science:

Math and science datasets. I will only put the most noteworthy ones here. If you need a lot of math data, check the midtraining data which will be sufficient.

SFT - Code/Agentic:

SFT - Misc:

SFT - Writing:

NSFW

Non-NSFW

Honestly, all of the above suck.

Preferences/Judging:

RL Non-Agentic:

RL Agentic:

TBD/Needs Review:

About

A collection of high quality huggingface datasets.

Resources

Stars

29 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors