Datasetpublic

fxmeng/UltraData-SFT-2605-no-think-32k-200k

UltraData-SFT-2605 · no_think · 32k–200k A length-filtered subset of the no_think split of openbmb/UltraData-SFT-2605, containing conversations whose token length falls in the 32k–200k range. This is the long-context tier intended for extended-context SFT. Two companion tiers were produced from the same source: Dataset Length range Records fxmeng/UltraData-SFT-2605-no-think-8k-32k 8k–32k tokens 623,421 this repo — fxmeng/UltraData-SFT-2605-no-think-32k-200k 32k–200k… See the full description on the dataset page: https://huggingface.co/datasets/fxmeng/UltraData-SFT-2605-no-think-32k-200k.

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes732downloads

Dataset Card

UltraData-SFT-2605 · no_think · 32k–200k

A length-filtered subset of the no_think split of openbmb/UltraData-SFT-2605, containing conversations whose token length falls in the 32k–200k range. This is the long-context tier intended for extended-context SFT.

Two companion tiers were produced from the same source:

Dataset	Length range	Records
`fxmeng/UltraData-SFT-2605-no-think-8k-32k`	8k–32k tokens	623,421
this repo — `fxmeng/UltraData-SFT-2605-no-think-32k-200k`	32k–200k tokens	8,784

Conversations shorter than 8k tokens were dropped; conversations longer than 200k tokens (15 in total) were also excluded.

Composition

Category	Records	Files
Code	550	255
IF	11	5
Math	952	229
Multi-lang-Knowledge	7,245	50
Multi-lang-Math	26	3
Total	8,784	542

Note: Chinese-general and Knowledge produced no conversations in the 32k–200k range and are therefore absent from this tier. The bulk of long-context records come from Multi-lang-Knowledge.

Total size: ~1.4 GB (uncompressed JSONL).

Format

Each line is one JSON object (JSONL):

json

{
  "uid": "IF_no_think_0001487",
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "source": "UltraData-sft-2605",
  "domain": "IF",
  "think_type": "no_think"
}

Field	Type	Description
`uid`	string	Unique record id (`<category>_no_think_<n>`)
`messages`	list	Chat turns, each `{role, content}`; `role ∈ {user, assistant}`
`source`	string	Always `UltraData-sft-2605`
`domain`	string	Category (Code / Math / IF / …)
`think_type`	string	Always `no_think` (no chain-of-thought traces)

Files are laid out one directory per category, sharded as <Category>_no_think_part-XXX-of-NNN.jsonl.

Usage

python

from datasets import load_dataset

# entire tier
ds = load_dataset("fxmeng/UltraData-SFT-2605-no-think-32k-200k", split="train")

# a single category
mlk = load_dataset("fxmeng/UltraData-SFT-2605-no-think-32k-200k", "Multi-lang-Knowledge", split="train")

Provenance & processing

1.Source: the no_think split of openbmb/UltraData-SFT-2605 (855 JSONL shards).
2.Each conversation's token length was measured and bucketed: < 8k → dropped, 8k–32k → companion repo, 32k–200k → this repo, > 200k → dropped.
3.Category/shard structure and record schema are preserved unchanged from the source.

License

Released under Apache-2.0, following the upstream openbmb/UltraData-SFT-2605 license. Please also cite/credit the original OpenBMB dataset.