Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
0
2.18k
diff --git a/moto/logs/models.py b/moto/logs/models.py
--- a/moto/logs/models.py
+++ b/moto/logs/models.py
@@ -1,5 +1,5 @@
from datetime import datetime, timedelta
-from typing import Any, Dict, List, Tuple, Optional
+from typing import Any, Dict, Iterable, List, Tuple, Optional
from moto.core import BaseBackend, BackendDict, BaseModel
from moto.core import CloudFormationModel
from moto.core.utils import unix_time_millis
@@ -85,22 +85,20 @@ def to_response_dict(self) -> Dict[str, Any]:
class LogStream(BaseModel):
_log_ids = 0
- def __init__(self, account_id: str, region: str, log_group: str, name: str):
- self.account_id = account_id
- self.region = region
- self.arn = f"arn:aws:logs:{region}:{account_id}:log-group:{log_group}:log-stream:{name}"
+ def __init__(self, log_group: "LogGroup", name: str):
+ self.account_id = log_group.account_id
+ self.region = log_group.region
+ self.log_group = log_group
+ self.arn = f"arn:aws:logs:{self.region}:{self.account_id}:log-group:{log_group.name}:log-stream:{name}"
self.creation_time = int(unix_time_millis())
self.first_event_timestamp = None
self.last_event_timestamp = None
self.last_ingestion_time: Optional[int] = None
self.log_stream_name = name
self.stored_bytes = 0
- self.upload_sequence_token = (
- 0 # I'm guessing this is token needed for sequenceToken by put_events
- )
+ # I'm guessing this is token needed for sequenceToken by put_events
+ self.upload_sequence_token = 0
self.events: List[LogEvent] = []
- self.destination_arn: Optional[str] = None
- self.filter_name: Optional[str] = None
self.__class__._log_ids += 1
@@ -133,12 +131,7 @@ def to_describe_dict(self) -> Dict[str, Any]:
res.update(rest)
return res
- def put_log_events(
- self,
- log_group_name: str,
- log_stream_name: str,
- log_events: List[Dict[str, Any]],
- ) -> str:
+ def put_log_events(self, log_events: List[Dict[str, Any]]) -> str:
# TODO: ensure sequence_token
# TODO: to be thread safe this would need a lock
self.last_ingestion_time = int(unix_time_millis())
@@ -152,9 +145,9 @@ def put_log_events(
self.events += events
self.upload_sequence_token += 1
- service = None
- if self.destination_arn:
- service = self.destination_arn.split(":")[2]
+ for subscription_filter in self.log_group.subscription_filters.values():
+
+ service = subscription_filter.destination_arn.split(":")[2]
formatted_log_events = [
{
"id": event.event_id,
@@ -163,41 +156,54 @@ def put_log_events(
}
for event in events
]
+ self._send_log_events(
+ service=service,
+ destination_arn=subscription_filter.destination_arn,
+ filter_name=subscription_filter.name,
+ log_events=formatted_log_events,
+ )
+ return f"{self.upload_sequence_token:056d}"
+
+ def _send_log_events(
+ self,
+ service: str,
+ destination_arn: str,
+ filter_name: str,
+ log_events: List[Dict[str, Any]],
+ ) -> None:
if service == "lambda":
from moto.awslambda import lambda_backends # due to circular dependency
lambda_backends[self.account_id][self.region].send_log_event(
- self.destination_arn,
- self.filter_name,
- log_group_name,
- log_stream_name,
- formatted_log_events,
+ destination_arn,
+ filter_name,
+ self.log_group.name,
+ self.log_stream_name,
End of preview. Expand in Data Studio

MS release evaluation texts

These are the texts used to measure the MS (Mad Shift) GGUF releases of Qwen3.8-27B and Qwen3.5-2B, and the unsloth GGUF files they are compared with: four English texts for both models, and one Chinese text for both models. The results are in the model cards (the Chinese results for Qwen3.8-27B are in the technical report): devvexus/Qwen3.8-27B-MS and devvexus/Qwen3.5-2B-MS.

These texts were not used to make the MS files. The MS converter never read them, and no calibration step used them. They were used to measure finished files. The Chinese text was used only for the Chinese measurements of both models, and it was fetched on 2026-11-16, after the MS files were made.

The wiki, web, code and Chinese texts are published exactly as they were used. The math text is not included (see Licenses); build_math.py rebuilds it exactly from its public source and checks it against the hashes below.

Files

File Bytes SHA-256 Source License
eval_large_wiki.txt 301,531 a38cbaac6a1abc8f95e89b03eb6493c5f2ec98c268b8a08210266f46227699bd 53 English Wikipedia articles (wikimedia/wikipedia, 20231101.en) CC BY-SA 4.0
eval_large_web.txt 340,670 c041ae5b573fef19d875b1f82596dc92294240cccac528faa35ef37d0aaf11fd 93 web pages (HuggingFaceFW/fineweb-edu, sample-10BT) ODC-By 1.0 + Common Crawl Terms of Use; four pages CC BY-SA
eval_large_code.txt 283,340 0ac3c7c2df8486320e9f8fc50ab7c8c0301ecbec76b414868345e83a5be6650a 32 code patches (SWE-Gym/SWE-Gym), all from github.com/getmoto/moto SWE-Gym: MIT; patches: Apache-2.0 (moto)
eval_large_math.txt (not included) 175,757 fda6fcf6ec025c8f59b2b0b8d76c7893a0429074a2d6c8c76a0a126294ed4861 222 MATH test problems with solutions (EleutherAI/hendrycks_math) Not included; rebuild with build_math.py
eval_large_zh.txt 394,771 75ab804edf1d995ec9ad28dbd68b29c68bb4586c7fe46438264191e9841c8f16 44 random articles from Chinese Wikipedia (zh.wikipedia.org, fetched 2026-11-16) CC BY-SA 4.0
build_math.py Our script: rebuilds eval_large_math.txt and checks its SHA-256 MIT
fetch_zhwiki.py Our script: fetches the random Chinese Wikipedia articles (how eval_large_zh.txt was built) MIT
LICENSE-moto.txt The Apache License 2.0 text from github.com/getmoto/moto Apache-2.0
provenance.json Row IDs, titles, URLs and byte offsets of every document Same as the text each entry describes
provenance_zh.json Title, page ID, revision ID, URL and character offset of every Chinese article Same as the text it describes

Each text is plain UTF-8. Documents are joined with one blank line. The English files have Windows line endings (CRLF), because they were written on Windows; eval_large_zh.txt has LF line endings. SHA-256 of the English files after converting CRLF to LF:

  • wiki 0e38421e680df1f84b81165ee0b2edc66e4c2d5b41c7416b4a131546ff739fda
  • web 00a522791b28739fca1a34cdbf658df7e664155968d52b90e41fd444021b9e6d
  • code c0e1c88fbe1d37b00ac31f6851dc1835e253f233d4223776a9a53d573a06c08c
  • math 0b4769ffdf081ba04949e6367b7b9f10e6c6e41c630c23521137527df92ba09e

To rebuild the math text (Python 3, standard library only; it reads the rows from the Hugging Face datasets-server API):

python build_math.py              # writes eval_large_math.txt with CRLF line endings, as evaluated
python build_math.py --lf         # LF line endings
python build_math.py --datasets   # reads the pinned revision with the `datasets` library instead

The script writes the file only if both SHA-256 hashes match.

How the texts were used

For each model and each text, a reference was made once from the BF16 model. Every GGUF file was then compared with that reference using llama-perplexity --kl-divergence:

# reference, once per text (BF16 model)
llama-perplexity -m Qwen3.8-27B-BF16.gguf -f eval_large_wiki.txt -c 512 -b 512 -ub 512 --chunks 128 -ngl 99 --kl-divergence-base wiki.kld

# each GGUF file
llama-perplexity -m <file>.gguf -f eval_large_wiki.txt -c 512 -b 512 -ub 512 --chunks 128 -ngl 99 --kl-divergence-base wiki.kld --kl-divergence

The same commands were run for web, code and math, and with Qwen3.5-2B-BF16.gguf for the 2B. MS files ran on ms-llama.cpp. The BF16 references and all other files ran on upstream llama.cpp 36a0e2ffc.

  • Each text is cut into 128 chunks of 512 tokens, so the first 65,536 tokens of each file are used. Both models' tokenizers give the same tokens for these files: 70,279 (wiki), 70,890 (web), 70,094 (code) and 70,073 (math). The last 4,500 to 5,400 tokens of each file are not used.
  • In each chunk, the second half is scored: positions 256 to 510 each predict the next token, so 255 tokens per chunk and 32,640 scored tokens per text.
  • With --kl-divergence, llama-perplexity reads the tokens from the reference file, so every GGUF file is scored on exactly the same tokens.
  • Chinese (both models): the same commands with -f eval_large_zh.txt and Qwen3.5-2B-BF16.gguf (2B) or Qwen3.8-27B-BF16.gguf (27B), against the same unsloth files and on the same builds. The file has 104,500 tokens; the first 65,536 are used (128 chunks, up to character 96,168 of 152,724), so it is also scored on 32,640 tokens.
  • The release runs were on Windows. There, llama-perplexity reads the file in text mode, which turns CRLF into LF. It does not process escape sequences in the file. On Linux or macOS, convert the line endings first (for example tr -d '\r' < eval_large_wiki.txt > wiki.txt) to get the same tokens. The LF hashes are listed above.

How the texts were built

The English texts were built on 2026-11-09 from the Hugging Face datasets-server rows API. Rows were read in order from a fixed starting row, and documents were added until a text held at least 70,000 tokens (Qwen3.8 tokenizer). A document was skipped if its first 200 characters already appeared in an older evaluation text. Every document's source row is in provenance.json. On 2026-11-16 the same rows were read again, and they rebuilt all four files byte for byte.

wiki. wikimedia/wikipedia, revision b04c8d1ceb2f5cd4588862100d08de323dccfbaa, config 20231101.en (the 1 November 2023 Wikipedia dump), split train, column text. Rows 6,300,000 to 6,300,052, all 53 used in order. The article titles and links are listed under Licenses.

web. HuggingFaceFW/fineweb-edu, revision 87f09149ef4734204d70ed1d046ddc9ca3f2b8f9, config sample-10BT, split train, column text. Rows 9,000,000 to 9,000,092, all 93 used in order. All come from Common Crawl dump CC-MAIN-2019-43. Each page's FineWeb-Edu id (a WARC record ID) and URL are in provenance.json.

code. SWE-Gym/SWE-Gym, revision bb94ed9e39bbeb96a7fcbfb533b80f25a7fd59cb, split train, column patch (the reference fix of each task, as a unified diff). Rows from 200 whose patch is longer than 2,000 and shorter than 30,000 characters: 32 patches from rows 201 to 267. All 32 are from the repository getmoto/moto, instances getmoto__moto- 6724, 4977, 7608, 6785, 4848, 5444, 5433, 5439, 7647, 6586, 7270, 7393, 6028, 5576, 5084, 5893, 4951, 6332, 5213, 5654, 6919, 6439, 6444, 6637, 5841, 5482, 6867, 4817, 7566, 6633, 6743 and 6734. The base commit of each is in provenance.json.

math. EleutherAI/hendrycks_math, revision 21a5633873b6a120296cce3e2df9d5550074f4a3, the test split of the seven subjects, taken in turn (algebra, counting and probability, geometry, intermediate algebra, number theory, prealgebra, precalculus), rows 50 to 81 of each. Two rows were skipped because they were in an older evaluation text (prealgebra row 52, counting and probability row 74), so 222 problems. Each document is problem + "\n\nSolution: " + solution. This copy has no problem IDs, so documents are identified by subject and row.

The revisions above are the datasets' main branches. None of them changed between their last update and 2026-11-09, so these are the revisions that were read.

zh. Chinese Wikipedia through its API (zh.wikipedia.org/w/api.php), fetched 2026-11-16 09:56-10:12 UTC with fetch_zhwiki.py: random main-namespace articles (generator=random), each article's plain-text extract (prop=extracts, explaintext), with its revision ID recorded. Skipped: disambiguation pages, list, index and timeline pages (by title), and articles whose plain text was under 2,000 or over 10,000 characters. Trailing reference, link and see-also sections were cut and runs of blank lines collapsed. The text is as stored on zh.wikipedia (simplified and traditional characters as the editors wrote them; no conversion). 44 articles were kept (513 random pages skipped), joined with one blank line in the order they were fetched; each article's title, page ID, revision ID and character offset are in provenance_zh.json.

Licenses

This dataset has no single license. Each file keeps the license of its source.

eval_large_wiki.txt: CC BY-SA 4.0. Text from English Wikipedia by Wikipedia contributors, under the Creative Commons Attribution-ShareAlike 4.0 License. Wikipedia text is also available under the GFDL. Changes: the plain text as prepared by wikimedia/wikipedia (markup removed), articles joined with blank lines, CRLF line endings. This file stays under CC BY-SA 4.0. The articles, in file order:

Galwegian; Villa Welgelegen; Berith; Gallovidian; MV Tustumena; Christy Smith (Survivor contestant); Battle of Mondovì; Second Battle of Tripoli Harbor; American Crime (film); Monika Kaelin; Gleason grading system; Viktor Bunyakovsky; Wacho; Battle of Aussig; Dog meat; Vanguard of Red Youth; Daniel Meyer (engineer); Deva Raya II; Audoin; German Naval Laws; Roseworthy College; Comparison of the imperial and US customary measurement systems; Patriot Guard Riders; Special Herbs, Vols. 7 & 8; Shadia Simmons; More Human Heart; Godepert; Evergreen International Aviation; Capitol Christmas Tree; Line–plane intersection; 58th Academy Awards; Jerry Alfred; Henry Ward Poole; Multinational Force in Lebanon; Manto (poem); Mallikarjuna Raya; Elena Ralph; John Tipton; Professor (disambiguation); Fakebook; Mojo Pin; Gary Clark (American football); Wendy Liebman; IAsiaWorks; Syed Abu Nasar; Fireside (LDS Church); William Harold Lee; Composant; Mobile phones and driving safety; Goncalo alves; Green libertarianism; Gesang Martohartono; Fredericton Express.

eval_large_zh.txt: CC BY-SA 4.0. Text from Chinese Wikipedia by Wikipedia contributors, under the Creative Commons Attribution-ShareAlike 4.0 License. Wikipedia text is also available under the GFDL. Changes: the plain-text extracts returned by the MediaWiki API (markup removed), trailing reference, link and see-also sections removed, runs of blank lines collapsed, articles joined with blank lines. This file stays under CC BY-SA 4.0. The articles, in file order, each linked to the revision used:

赤鱲角; 雅夫; 賴昌星; 陳嘉上; ElDLIVE宇宙警探; 安岡章太郎; 托馬斯·林賽 (學者); 地狱 (小说); 梁舜燕; 圣维特主教座堂; 误差函数; 乳糖不耐; 2002年热带风暴费伊; 谷口悟朗; 三橘愛; 科尼岛 (歌曲); 3月14日; 限購令; 变应性结膜炎; 乌里扬诺夫斯克汽车制造厂; 高松信司; 广州恒大足球俱乐部2014赛季; 本·巴雷斯; 施塔恩贝格湖; 小花 (小魔女DoReMi); 沃伊洛维察; 攻擊者級護航航空母艦; 2024年巴基斯坦大选; 古弗尼尔·莫里斯; 遙控伺服; JR東日本E259系電力動車組; 农麦青山; RCA Studio II; 南瓜酒; 桑德贝; 民主300+; 楊儒賓; 山原秧雞; 尼尼微人以撒; 紅烏龜:小島物語; 塞尔达传说 旷野之息; 葡萄牙语正写法; 吴庆华; 托尼·路特曼.

eval_large_web.txt: ODC-By 1.0 and the Common Crawl Terms of Use. FineWeb-Edu is released under the Open Data Commons Attribution License (ODC-By) v1.0, and its use is also subject to Common Crawl's Terms of Use. The pages themselves remain subject to their publishers' rights; each page's URL is in provenance.json. Four pages carry their own license:

Row Page License
9,000,057 Bunsen burner, Simple English Wikipedia contributors (as crawled in October 2019) CC BY-SA 3.0, also GFDL
9,000,069 Why is American and French notation different for open intervals (x, y) vs. ]x, y[?, History of Science and Mathematics Stack Exchange; question by Franck Dernoncourt, answer by Alexandre Eremenko CC BY-SA 3.0
9,000,077 match-fixing tag info, Sports Stack Exchange tag wiki, last edited by Nij (earlier editors in the page history) CC BY-SA 4.0
9,000,079 What is the ideal time period for 'forgotten history' novels/stories to take place and still end up with our history?, Worldbuilding Stack Exchange; the crawled page text is one answer, by zagdrob CC BY-SA 3.0

eval_large_code.txt: MIT (SWE-Gym) and Apache-2.0 (moto). The SWE-Gym dataset is released under the MIT license (verified on its dataset card). The patches are unchanged code from getmoto/moto, licensed under the Apache License 2.0; the license text is in LICENSE-moto.txt. moto has no NOTICE file.

eval_large_math.txt: not included. The math text is not included in this dataset. In January 2025 Art of Problem Solving filed a copyright (DMCA) notice against the MATH problems on Hugging Face, and the Hugging Face copy of the MATH dataset (hendrycks/competition_math) was taken down. Because of that claim, we do not redistribute the problems. Instead we publish exactly which problems were used (in provenance.json) and a script that rebuilds the identical file from the EleutherAI/hendrycks_math copy and checks its hash. We would publish the text itself if we could. We claim no license for the math text. build_math.py is ours, under the MIT license.

Personal data

The texts were checked for personal data, keys and tokens before publication. No private personal data and no real keys or tokens were found. People are named as they appear in Wikipedia articles, news pages and organization pages. The web text is FineWeb-Edu as served, and includes a few low-quality pages.

Not in this dataset

The MS converter and its calibration data are not part of this dataset.

Citations

Wikipedia:

@ONLINE{wikidump,
    author = "Wikimedia Foundation",
    title  = "Wikimedia Downloads",
    url    = "https://dumps.wikimedia.org"
}

FineWeb-Edu (paper: arXiv:2406.17557):

@misc{lozhkov2024fineweb-edu,
    author       = { Lozhkov, Anton and Ben Allal, Loubna and von Werra, Leandro and Wolf, Thomas },
    title        = { FineWeb-Edu: the Finest Collection of Educational Content },
    year         = 2024,
    url          = { https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu },
    doi          = { 10.57967/hf/2497 },
    publisher    = { Hugging Face }
}

SWE-Gym:

@inproceedings{pan2025training,
  title     = {Training Software Engineering Agents and Verifiers with SWE-Gym},
  author    = {Pan, Jiayi and Wang, Xingyao and Neubig, Graham and Jaitly, Navdeep and Ji, Heng and Suhr, Alane and Zhang, Yizhe},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML 2025)},
  year      = {2025},
  note      = {arXiv:2412.21139, accepted at ICML 2025},
  url       = {https://arxiv.org/abs/2412.21139},
}

MATH:

@article{hendrycksmath2021,
  title={Measuring Mathematical Problem Solving With the MATH Dataset},
  author={Dan Hendrycks and Collin Burns and Saurav Kadavath and Akul Arora and Steven Basart and Eric Tang and Dawn Song and Jacob Steinhardt},
  journal={NeurIPS},
  year={2021}
}

The code patches come from moto: https://github.com/getmoto/moto.

Downloads last month
24

Collection including devvexus/ms-eval-texts

Papers for devvexus/ms-eval-texts