THUDM / THUDM/slime

about large data to train

Open
#1,231 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
8.5k
Forks
1.3k
Avg merge
5h 36m
Merged PRs (30d)
22

Description

https://github.com/THUDM/slime/blob/ab880fc8cf5b644011f53b2e4b4e5c405d174873/scripts/run-qwen3-4B-base-sft.sh#L39
when I try to conduct SFT on large data, the process stuck on data load. There are two main problems:
1.why the prompt-data must be a single json or parquet file? Sometime we need to SFT on a large data which may need to be splited to many parts. The framework should support using a file dir including multiple small parts.
2.when I use a single file about 350G containing 15M samples, the process will stuck on dataload and finally failed.
Overall, I have a need to SFT on about 15M samples and how can I achieve this? I cannot use a file dir containing multiple data parts, neither can use single large dataset? Thanks!

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start at scripts/run-qwen3-4B-base-sft.sh#L39 and trace how the prompt data path is loaded. Reproduce the reported behavior with a multi-part directory and a roughly 350G dataset, then verify that large SFT datasets can load without stalling and that the requested input forms are supported.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
data, machine-learning
Issue type
Feature
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.