OfflineBase/Muse-Glimmer-30B-2026-Knowledge-Data

Muse Glimmer 2026 Knowledge: the stage-b-r30b training and eval data

Built with Muse Spark Open.

The model trained on these files is OfflineBase/Muse-Glimmer-30B-2026-Knowledge (stage-b-r30b). Rows written by Muse Spark Open or by our tuned model are under the Muse Model License: read "License for Spark-written and model-output rows" before you use them.

Built Oct 31, 2026. Counts come from the build's release report, which stays with the build files. File sizes and sha256s are those of the files in this repository.

What this is

Chat-format question and answer pairs about events from January 4 to October 29, 2026, plus the Stage B honesty, reasoning and measured rows and every evaluation set used to train and score stage-b-r30b (OfflineBase/Muse-Glimmer-30B-2026-Knowledge), a LoRA fine-tune of meta-models/Muse-Glimmer-30B. The owner's scope (Oct 31, 2026) is everything: all three training files, their components, the tuned model's measured answers, the eval sets, news-derived rows and Muse Spark Open-written rows.

880,648 input rows went in (the same rows repeat across files, as they did in training). 649,133 rows are published, in six license groups (each file is its own config; see "Loading"). Rows were removed only for the reasons in "What was removed": personal data about private individuals, source licenses that forbid redistributing derived works, court records, verbatim overlap with publisher text and NC sources, publisher rows the verbatim check could not run on, long passages copied from open sources, unknown writers, and Exa or LittleBit markers. No source passages or article text are included. A second reader re-checked the build and removed 3,251 more rows (see "Second read"), and a last pass removed the 3,686 news rows whose source records were missing, so the verbatim check never ran on them (see "Verbatim guard").

Size: 649,133 rows and 46,692,158 tokens of text (about 46.7 million), counted with the model's own tokenizer. In the chat format the model trains on, the chat rows come to 71,948,804 tokens, about 74.8 million with the other rows' text added. Per config: see "Tokens".

Configs and licenses

Config	License	Rows	Compressed bytes
public-sources	CC BY-SA 4.0 (127,369 rows); CC BY-SA 3.0 for the 36,407 Dolly-derived rows	163,776	40,472,305
news-derived	News-derived note (below); no license is granted in publisher text	312,164	38,783,201
noncommercial	CC BY-NC-SA 4.0; 557 of these rows are also under the Muse Model License	952	84,746
spark	Muse Model License, "Built with Muse Spark Open"; each row also CC BY-SA 4.0 (85,358) or the news-derived note (33,216)	118,574	19,791,074
model-outputs	Muse Model License, "Built with Muse Spark Open"; each row also CC BY-SA 4.0 (12,365) or the news-derived note (36,462)	48,827	6,204,911
evals	Per row, in release.row_license: CC BY-SA 4.0 (3,148) or the news-derived note (1,692); 2,893 of these rows are also under the Muse Model License	4,840	517,215

Tokens

Counted with the stage-b-r30b tokenizer (the same as Muse Glimmer 30B's). Text tokens count the message contents, or the question, prompt, premise, answer and reply fields for rows without messages. Chat-format tokens count chat rows rendered with the model's chat template, as in training. The evals and measured rows have no messages, so they have no chat-format count.

Config	Rows	Text tokens	Chat-format tokens
public-sources	163,776	16,680,148	29,708,820
news-derived	312,164	17,876,595	27,853,062
noncommercial	952	54,910	80,628
spark	118,574	9,737,239	13,530,881
model-outputs	48,827	2,213,999	775,413
evals	4,840	129,267	-
All	649,133	46,692,158	71,948,804

Files, one per split:

Group	Split	Config to load	File	Rows	Bytes	sha256
evals	cutoff_20	evals-cutoff_20	evals/cutoff_20.jsonl.gz	20	897	17f9fb8c538a355aaaec808b80ce60fd4d5e7c48d1e52ee76c7e86b7db275606
evals	honesty_190	evals-honesty_190	evals/honesty_190.jsonl.gz	189	17,992	b902da772b1830916620b9ef68d8b1c023e9fe97d13d3ffc6e58e5c67169f7f5
evals	honesty_dev_120	evals-honesty_dev_120	evals/honesty_dev_120.jsonl.gz	81	13,554	f8c81345c834d2e6b7848024dbd4da87656a8b3ee4ac0d4a672e7b406eca0257
evals	named_controls_150	evals-named_controls_150	evals/named_controls_150.jsonl.gz	148	16,219	12fd1bbadc26ef9f56435557e7993fbdbed1306d640689014c1909fc01717231
evals	old_facts_40	evals-old_facts_40	evals/old_facts_40.jsonl.gz	40	1,801	ebbcc6f6e4f8a9c6cd507a39ed2fe0a2ad23af962801a9ca4fc45b9df78489e5
evals	probes_1500	evals-probes_1500	evals/probes_1500.jsonl.gz	1,433	168,887	f922578e2805f3fd8b4d98ad6d4b57734b230c212175d3a4610d596c330e5cc6
evals	probes_2500	evals-probes_2500	evals/probes_2500.jsonl.gz	2,426	266,153	cef6dd14dfe12acaf16e4802a37515435015145b7ad22c8270e13d9e163e51be
evals	reasoning_260	evals-reasoning_260	evals/reasoning_260.jsonl.gz	260	15,251	cb13c59b43deead9fc8f943ea5af77be7466eddac6ee8eedeee080e92c2026b1
evals	salient_246	evals-salient_246	evals/salient_246.jsonl.gz	243	16,461	cf91bdc6ce352fcd6502d7b6419634de5610b81dd14d8b5316df4b557e102dbd
model-outputs	measured	model-outputs-measured	model-outputs/measured.jsonl.gz	38,486	5,148,436	4cb886b083c11092520997e59fe7e668b733b5fddb048115e490a3d9af2838ca
model-outputs	stage_b_r30b	model-outputs-stage_b_r30b	model-outputs/stage_b_r30b.jsonl.gz	10,341	1,056,475	10ab3db397e65773c2a3d5f34b4ccec35efe54de82c82763efe17f1bd80b1451
news-derived	stage1	news-derived-stage1	news-derived/stage1.jsonl.gz	3,880	444,330	9a62edc38c6e622e72b6a7e50da8d98da657d8b386d2f68316419ebc724bc953
news-derived	stage2	news-derived-stage2	news-derived/stage2.jsonl.gz	283,366	35,018,032	b763f8c5f11efa6808ccdcf052303747e87dae88c41a0eea3116640fd7abc1a8
news-derived	stage_b_r30b	news-derived-stage_b_r30b	news-derived/stage_b_r30b.jsonl.gz	15,610	2,014,671	54597c09f9c816743c28f0298363d4c604c3fdd3b16e64981831d19e23bc4369
news-derived	stage_b_written	news-derived-stage_b_written	news-derived/stage_b_written.jsonl.gz	9,308	1,306,168	024481fedfaea65549506565b97da6b5d6ace286ec40b22f76663b7eb4932818
noncommercial	honesty_dev_120	noncommercial-honesty_dev_120	noncommercial/honesty_dev_120.jsonl.gz	1	642	b2c6f8f356c069d3d83a41af7bc91e8aab0aaa6badb45c52b5991214dff5ea1b
noncommercial	measured	noncommercial-measured	noncommercial/measured.jsonl.gz	63	6,775	08506cda733049786d310cc21c1e46b5232cf7d6d6a23eb1cff542de35bc11c1
noncommercial	stage1	noncommercial-stage1	noncommercial/stage1.jsonl.gz	472	45,356	d4c132f389033dd6759f99cb48099f7db61c9a96c34198a89e53fffff40849f6
noncommercial	stage2	noncommercial-stage2	noncommercial/stage2.jsonl.gz	322	21,327	3ec7aeb66c0ddb3f1b9e336ea1d5b9fe20c3f11a4c94e0c7b17eb29c2d136077
noncommercial	stage_b_r30b	noncommercial-stage_b_r30b	noncommercial/stage_b_r30b.jsonl.gz	50	4,996	f7d50df6cc1a0eea03e5afb974ed80180fdc916eaa35cbbf38f40c52e429f814
noncommercial	stage_b_written	noncommercial-stage_b_written	noncommercial/stage_b_written.jsonl.gz	44	5,650	af562a53e65fd497e57db39bdf0c196f99bce4febeb7df9caea957a9f8b8db4d
public-sources	replay	public-sources-replay	public-sources/replay.jsonl.gz	6,519	1,756,796	6afac9051a1bbfc2292203ac4064061a43a3199fde03791b00d71a3a98baac62
public-sources	stage1	public-sources-stage1	public-sources/stage1.jsonl.gz	41,362	6,636,554	9c975a800934dd1d2ddbac9feac46771dbde6fa643693e2f6a1017f05e00d486
public-sources	stage2	public-sources-stage2	public-sources/stage2.jsonl.gz	58,075	9,218,884	9571e7330ca534cc1ea49b45cd9410015a0f7ec2b0e4f098d8f0053ab8246904
public-sources	stage_b_r30b	public-sources-stage_b_r30b	public-sources/stage_b_r30b.jsonl.gz	25,261	9,947,307	d76c21526e33e7c3d54d7aee8c51359a1ff9ac7515133bfc2b00701ecc559b3b
public-sources	stage_b_reasoning	public-sources-stage_b_reasoning	public-sources/stage_b_reasoning.jsonl.gz	26,336	11,970,748	defcb3080c37c323ddd21da35fbab6030ac5e87457bd62449e31caabfca66dc1
public-sources	stage_b_written	public-sources-stage_b_written	public-sources/stage_b_written.jsonl.gz	6,223	942,016	996addd085c66afc3fb22d0de495ac4fb2eabb17dd58d26d86df3f9ab12ad96c
spark	anchors	spark-anchors	spark/anchors.jsonl.gz	3,128	244,886	360201be2de25c9144eab7086e739acbf1ce96354338954441feb8570db34561
spark	stage1	spark-stage1	spark/stage1.jsonl.gz	112,318	18,125,788	f5e016c31f6b875f6cdcf978d9e73191add0643af5f87e5d6a4d9992b7a51f02
spark	stage2	spark-stage2	spark/stage2.jsonl.gz	3,128	1,420,400	9b3e03ffe4a5ee4c1e961507f13713219ca07020e3adc9f020a1c374b95934a6

Loading

Each file is its own config, named <group>-<split>, with one split, train. Files in the same license group have different columns (for example, public-sources/replay and model-outputs/measured have no messages list, and each eval set has its own fields), so they cannot be loaded as splits of one dataset. Load one file at a time:

from datasets import load_dataset

probes = load_dataset("OfflineBase/Muse-Glimmer-30B-2026-Knowledge-Data", "evals-probes_1500", split="train")

Elsewhere in this card, "config" means the license group (release.config), not the config you load.

License files in this repository: MUSE_MODEL_LICENSE.md (the Muse Model License, with the Muse Spark Open Usage Policy as Exhibit A) and USAGE_POLICY.md (the Muse Glimmer Usage Policy).

How rows were assigned, first match wins:

1. The source record's license is non-commercial (CC BY-NC, BY-NC-SA, cc-by-nc-4.0, FAIR noncommercial research): noncommercial, whatever the file. NC rows appear nowhere else.
2. The row is from an eval or test set: evals.
3. The row is the tuned model's own answer (Stage B measured rows and the 40,000 measured answers): model-outputs.
4. Muse Spark Open wrote the row (any host: Fireworks, Baseten, Modal), anchors included: spark.
5. The source is publisher text (CC-NEWS and other records our collection marked training only): news-derived.
6. Everything else: public-sources (Wikipedia, Dolly replay, government and openly licensed sources, templates, base Muse Glimmer reasoning rows).

A row can carry more than one set of terms (for example a Spark Open row written from a CC-NEWS article). Every row says which apply in its own release object.

The release object on every row

Field	Meaning
config, split	Where the row is published: the license group and the split (the file is <config>/<split>.jsonl.gz).
from_file, line	The original training or eval file and 1-based line, for lineage.
source_file	The collected-source file the row's passage came from (wikipedia, ec-press.jsonl, s3-ccnews-general.jsonl, ...), or - for rows with no source passage.
source_license	The source's terms and credit line, as our collection recorded them. For publisher text, the news-derived note.
source_terms_recorded	Publisher rows only: the terms string our collection stored. "Training only" in it describes the original collection rule, which the owner lifted for this release.
writer	Who wrote the row. The values are muse-glimmer-30b, muse-glimmer-30b (base, replay), muse-spark-open@<host>, template, stage2-multi, template, labelled by stage2-multi (template replies chosen by the tuned model), stage2-multi (reply); question and gold answer: <writer> (measured answers, where <writer> is muse-glimmer-30b, muse-spark-open@<host> or muse-glimmer-30b or muse-spark-open@<host>), and written by the coding agent or a script (no model writer) (eval sets written in the repo).
row_license	The terms that govern this row in the release: one or more of CC BY-SA 4.0, CC BY-SA 3.0 (Dolly-derived rows), CC BY-NC-SA 4.0, the news-derived note, and the Muse Model License with "Built with Muse Spark Open".
credit	Wikipedia rows: article link, title and revision id (permalink uses oldid) when the source records have them. Current-events rows link the Portal:Current events day page; no revision was recorded for those.
tuned_model, correct	Model-output rows: the model that answered (stage2-multi) and whether the answer was graded right.
eval	Eval rows: {set, contaminated_for_future_runs: true}.
verbatim_guard	Publisher rows: passed, or unchecked_anchor for anchors, which have no single source to compare against.

Other row fields are as trained: messages (system, user and assistant turns; Stage B reasoning rows also carry reasoning_content), kind and lineage in Stage B, and each eval set's own fields. source_type names the kind of source and source_id is the first 16 hex characters of the sha256 of the original source string (unsalted, so the owner can join rows to internal lineage). Removed from every row: facts, lineage.server, source, passageId, url (article URLs for publisher rows) and visibility. One exception: every evals/salient_246 row keeps its top-level source_url, the public page the question was drawn from (243 rows: 201 English Wikipedia pages and 42 Meta Newsroom pages on about.fb.com). The set was written in the repo, and the verbatim guard did not run on it. The visibility field was a build-time flag ("private" meant training only) and said "private" on most kept rows, so it is gone; release.config replaces it.

License per config

In the metadata above, other stands for the Muse Model License (MUSE_MODEL_LICENSE.md) and for the news-derived note, which grants no license.

This is not an official Meta release and is not endorsed by Meta. The names Muse, Muse Glimmer and Muse Spark appear only to say what these rows and the model are built on.

• public-sources: CC BY-SA 4.0, and CC BY-SA 3.0 for Dolly-derived rows. Rows from English Wikipedia and the other open sources are CC BY-SA 4.0 (127,369 rows). The 36,407 Dolly-derived rows (a databricks-dolly-15k prompt with a reply by base Muse Glimmer 30B: the replay rows and the Stage B reasoning rows on Dolly prompts) are CC BY-SA 3.0, the Dolly license, and their row_license says so (replay 6,519, stage1 6,519, stage2 6,519, stage_b_r30b 9,100, stage_b_reasoning 7,750). The other sources are public domain US government works, CC BY 4.0 and CC BY 3.0 (Global Voices), Open Government Licence v3.0, Open Parliament Licence v3.0, Apache 2.0, MIT, BSD-2/BSD-3-Clause, CC0, and reuse-with-credit terms of the European Medicines Agency, the ECB, the Bank of Canada and the BIS. All allow redistribution with credit; the BIS terms also ask that no extra charge is made for its statistics. Keep each row's source_license credit line when you reuse it. Base Muse Glimmer 30B, which wrote most of these rows, is Apache 2.0.
• news-derived: plain note, no license granted in source text. The facts and Q&A in these rows were written by models (Muse Glimmer 30B) from publisher articles: CC-NEWS (Common Crawl), SEC 8-K filings, Meta Newsroom, Meta engineering and docs, PyTorch blog, UN News, Agencia Brasil, GitHub release notes, OpenAlex abstracts without an NC or ND license, GDACS alerts and similar. The source articles remain under their publishers' copyright and we make no claim to them. No article text is included, and rows that share 12 or more consecutive words with their source were removed (see "Verbatim guard"). Edge cases kept here under the same note: GDACS alerts (no reuse license stated), Launch Library 2 (asks not to forward data without adding value), Banco de Mexico pages (no republishing verdict), medRxiv items marked cc0_ng or none, and GitHub release notes of NOASSERTION or GPL repositories.
• noncommercial: CC BY-NC-SA 4.0. Rows whose source is NC licensed (PMC, bioRxiv, medRxiv and OpenAlex items marked CC BY-NC or BY-NC-SA, Hugging Face cards under cc-by-nc-4.0 or the FAIR Noncommercial Research License, and World Bank documents under its non-commercial general terms). No commercial use. The FAIR Noncommercial Research License is narrower than CC BY-NC-SA (research use only); rows from those cards say so in source_license. The 557 rows written here by Muse Spark Open or stage2-multi are also under the Muse Model License (see "License for Spark-written and model-output rows"). These sources are not openly licensed text, so the 12-word verbatim guard runs on them as it does on publisher text.
• spark: Muse Model License. Rows written by Muse Spark Open (hosted on Fireworks, Baseten and Modal). Built with Muse Spark Open. The owner chose to publish the Fireworks-hosted rows without waiting for Fireworks' answer on its terms 2.2(c). Each row also keeps its source's terms: CC BY-SA 4.0 for rows written from Wikipedia or other open sources (85,358), and the news-derived note for rows written from publisher text and for the anchors (33,216). What the Muse Model License asks of you is in "License for Spark-written and model-output rows".
• model-outputs: Muse Model License. The tuned stage2-multi model's own answers, plus 3,560 template replies in stage_b_r30b that were picked because stage2-multi answered wrong. stage2-multi was trained on Muse Spark Open outputs, so it is a Muse Model License Derivative and its outputs are treated as Built with Muse Spark Open. Each row also keeps its source's terms: CC BY-SA 4.0 (12,365) or the news-derived note (36,462).
• evals: per row. Eval rows keep their source's terms in release.row_license: CC BY-SA 4.0 for sets written in the repo and for probes from Wikipedia and other open sources (3,148), and the news-derived note for probes from publisher text (1,692). The 2,893 probes and controls written by Muse Spark Open are also under the Muse Model License.

License for Spark-written and model-output rows

170,851 rows were written by Muse Spark Open or by our tuned model stage2-multi, a Muse Model License Derivative (counting the 3,560 template replies its wrong answers picked). They are under the Muse Model License (MUSE_MODEL_LICENSE.md, with the Muse Spark Open Usage Policy as Exhibit A) as well as the other terms in their release.row_license. Counts are rows as published; some rows repeat across splits, as they did in training.

Config	Split	Rows under the Muse Model License	Also CC BY-SA 4.0	Also CC BY-NC-SA 4.0	Also the news-derived note
spark	anchors	3,128	0	0	3,128
spark	stage1	112,318	85,358	0	26,960
spark	stage2	3,128	0	0	3,128
model-outputs	measured	38,486	9,739	0	28,747
model-outputs	stage_b_r30b	10,341	2,626	0	7,715
evals	named_controls_150	25	21	0	4
evals	probes_1500	1,433	1,086	0	347
evals	probes_2500	1,435	1,087	0	348
noncommercial	measured	63	0	63	0
noncommercial	stage1	470	0	470	0
noncommercial	stage_b_r30b	24	0	24	0
Total		170,851	99,917	557	70,377

If you copy or use these rows:

• Keep MUSE_MODEL_LICENSE.md and USAGE_POLICY.md with any copy of them, in full or in part.
• Show "Built with Muse Spark Open" in the product that uses them, or in its documentation, website or settings.
• Follow the usage policies: the Muse Spark Open Usage Policy (Exhibit A of MUSE_MODEL_LICENSE.md) and the Muse Glimmer Usage Policy (USAGE_POLICY.md). Among other terms, the policies say the models are not intended for people under 18, and that outputs, these rows included, must not be presented as written by a person.
• A model trained, fine-tuned or distilled on these rows is a Derivative under the Muse Model License, and the license applies to it too, including its Section 3 on Model as a Service businesses above USD 20 million in yearly revenue.

This is a summary; the license files are the terms.

Disclosure: rows under two licenses. Some Spark-written rows were written from Wikipedia text and carry both CC BY-SA 4.0 and the Muse Model License: 34,050 rows (32,854 in spark/stage1, 1,196 in the eval probes and controls). Counting every row that carries both, there are 99,917: those 34,050, 53,502 Spark-written rows from other openly licensed sources, and 12,365 model-outputs rows (7,938 of them on questions from Wikipedia text). Another 557 noncommercial rows carry both CC BY-NC-SA 4.0 and the Muse Model License. Whether both licenses can apply to the same rows is not settled. We publish them with this disclosed, and readers who need certainty should get their own legal advice.

Lineage of stage-b-r30b

All three steps are LoRA rank 64, alpha 64, on every language projection, 1 epoch, effective batch 32, max length 2,048, on 8x A100 80GB (Unsloth 2026.10.9). Base weights: meta-models/Muse-Glimmer-30B, BF16 snapshot cec94fcfd651. The result is published as OfflineBase/Muse-Glimmer-30B-2026-Knowledge.

Order	Run	Training file	Rows	Starts from	lr
1	stage1-multi	train-stage1-frozen.jsonl	199,309	base	1e-4
2	stage2-multi	train-stage2.jsonl	520,895	stage1-multi adapter	1e-4
3	stage-b-r30b	train-stage-b-r30b-shuffled.jsonl	60,000	stage2-multi adapter	5e-5

• Stage 1: 188,170 fact pairs from 43,225 passages, plus 4,400 overviews ("anchors"), 6,739 replay answers and 1,185 cutoff templates. Writers: Spark Open 138,143 fact rows plus the 4,400 anchors; Muse Glimmer 48,842; base Muse Glimmer replay 6,739; templates 1,185.
• Stage 2: 509,756 fact pairs from 195,730 passages, no question that stage 1 trained on or a probe holds, bulky government and science sources capped at 3,000 each, plus the same anchors, replay and templates as stage 1 (12,324 rows repeat word for word). About 80% of rows come from CC-NEWS (s3-ccnews-general 216,600, s3-ccnews-ai-meta 200,312).
• Stage B r30b (mix b, 30% reasoning, seed 20261031): contrast 20,397 (stage 2 fact rows), reasoning 18,000 (base Muse Glimmer's own reasoning and answer), measured right 7,023 (stage2-multi's own right answers), replay 4,091, measured wrong 3,652 (a template reply, chosen because stage2-multi was wrong), false premise 3,023, after cutoff 1,621, unknowable 1,392, cutoff restated 801. 201 contrast rows were marked "Spark Open or unmatched" at assembly; the Q&A cache lists only Muse Glimmer for their passages, so they are attributed to Muse Glimmer here.
• Components published in full (after removals): anchors (4,400), replay (6,739), stage-b-written (17,360, the pool the Stage B false premise, after cutoff and unknowable rows came from), stage-b-reasoning (26,919, the reasoning pool) and measured (40,000 stage2-multi answers to held-out questions, with right).
• Run metadata (hyper.json, metrics, training logs) is not in this release.

Input files and their sha256:

File key	Path	Rows	sha256
stage1	.data/knowledge/dataset/train-stage1-frozen.jsonl	199,309	fee0b04444bed6e069b5cf2ffb3bdaac6d825c66f34e1de335e268df3511f9d5
stage2	.data/knowledge/dataset/train-stage2.jsonl	520,895	fc40310a00d2b793fad512b816bae3795f4eefa1ad8cc75ca28cb73bad0178dd
stage_b_r30b	.data/knowledge/dataset/train-stage-b-r30b-shuffled.jsonl	60,000	640dbecf4e31e99c24d527312a3be1492a1ad92ec371c235fbf24c41f2cd0234
anchors	.data/knowledge/dataset/anchors.jsonl	4,400	851c8244ac115ba9318c18a4dcb912f5c8fa118f793f16345c873aa89e55dc16
replay	.data/knowledge/replay.jsonl	6,739	1ea0426e37ebc3124bc137f538569437415d2470a3090e4b9a049d344072a505
stage_b_written	.data/knowledge/dataset/stage-b-written.jsonl	17,360	765b9ecad6cd36863ec49f9cd3cd81da7c2697cbc9ec2b2b155267b2953b0889
stage_b_reasoning	.data/knowledge/dataset/stage-b-reasoning.jsonl	26,919	5640e62810d2ec2999f03b117d92e020f74310012327efce9e1792b36f7ef2de
measured	.data/knowledge/runs/stage-b-measure/measured.jsonl	40,000	3841f94f024e4c74fd106e8b79bfec0db6110e0bc56b1ad519a667df5ce75611
eval_probes_1500	.data/knowledge/dataset/eval-probes.jsonl	1,500	899c48d4fcd4fe0b4f02b043361c4e69064ace72ee4b2b732e756427165f5924
eval_probes_2500	.data/knowledge/dataset/stage2-eval-probes.jsonl	2,500	6703716fc7fd6133cbd5c7cbd3501da26130fffbc93e6976e9046f9d6fea4b9b
eval_honesty_dev_120	.data/knowledge/dataset/stage-b-honesty-dev.jsonl	120	d5a3fb32783e9c0233e2c443dd61e80340ffe56efc43f27e58876ff146de3d5f
eval_named_controls_150	.data/knowledge/dataset/stage-b-named-controls.jsonl	150	0e2876ebbea5c75f5e74d97611465d1f0072fa33e49fca216eaaa4d509bc083a
eval_honesty_190	scripts/knowledge/aws/job/honesty_probes.jsonl	190	28a0287af0ef3cfddd0ae65e5b4b298ce28a95898b5cbaa30be72badcde479b8
eval_salient_246	scripts/knowledge/aws/job/salient_probes.jsonl	246	a53b574bdcbc42993e2c82cc3b68278fcca10e941004465e79ecbb4604ba73b8
eval_reasoning_260	scripts/knowledge/aws/job/reasoning_probes.jsonl	260	c7ce54a78a37a63ff7a485807b26eeb07760e462ab2544977542c1dbe58fe5a3
eval_old_facts_40	scripts/knowledge/aws/forget.jsonl	40	550535ad73ff9d7b29d2785df6253dc0dedd0f159c298bb5f78ed700b50deef4
eval_cutoff_20	scripts/knowledge/aws/job/cutoff_questions.jsonl	20	8b4a083689af683379013c3ff259e78a21093e3b1750a70ea1d4195ddc825d31

What was removed

Rows in, rows published and removals per input file. A row can hit several rules; it is counted once, under the first rule in this order: Exa or LittleBit marker, court record, NEVER list, no-reuse license, ND license, missing source record, unknown source, unknown writer, private-people list, system prompt, incident article, named killing or case article, hard personal-data rule, soft personal-data rule. The verbatim guard ran after these, then the second read, and last the removal of news rows the verbatim guard could not check ("Unchecked news").

File	In	Published	Personal data (hard / soft / article rules)	License (ND / no-reuse)	Court	Exa	Verbatim	Second read	Unchecked news
stage1	199,309	158,032	1,491 / 5,022 / 4,729	772 / 224	615	0	22,855	1,883	3,686
stage2	520,895	344,891	2,674 / 8,072 / 2,046	0 / 0	0	8	162,655	549	0
stage_b_r30b	60,000	51,262	424 / 571 / 168	24 / 11	0	1	7,347	192	0
anchors	4,400	3,128	376 / 852 / 0	0 / 0	0	0	0	44	0
replay	6,739	6,519	116 / 100 / 0	0 / 0	0	0	0	4	0
stage_b_written	17,360	15,575	45 / 241 / 87	65 / 29	0	0	795	523	0
stage_b_reasoning	26,919	26,336	336 / 239 / 0	0 / 0	0	2	0	6	0
measured	40,000	38,549	265 / 502 / 431	15 / 3	2	0	205	28	0
eval_probes_1500	1,500	1,433	11 / 9 / 32	0 / 4	2	0	1	8	0
eval_probes_2500	2,500	2,426	12 / 14 / 32	0 / 4	2	0	2	8	0
eval_honesty_dev_120	120	82	0 / 2 / 0	0 / 0	0	1	29	6	0
eval_named_controls_150	150	148	0 / 1 / 0	1 / 0	0	0	0	0	0
eval_honesty_190	190	189	0 / 1 / 0	0 / 0	0	0	0	0	0
eval_salient_246	246	243	0 / 3 / 0	0 / 0	0	0	0	0	0
eval_reasoning_260	260	260	0 / 0 / 0	0 / 0	0	0	0	0	0
eval_old_facts_40	40	40	0 / 0 / 0	0 / 0	0	0	0	0	0
eval_cutoff_20	20	20	0 / 0 / 0	0 / 0	0	0	0	0	0

Totals by first reason: pii_soft 15,629, incident_article 7,511, pii_hard 5,723, license_nd 877, court_record 621, license_no_reuse 275, private_people_list 27, exa_marker 12, incident_or_sensitive_article 10, named_killing_or_case_article 4; verbatim guard 193,889; second read 3,251; unchecked news 3,686. No row was removed for an unknown writer, an unknown source, a missing record of a non-news source, the NEVER list or a LittleBit marker (all 0). Exa markers are word matches only (no Exa result was ever stored); those rows were removed anyway.

Personal data

Personal data about private individuals is not published. The rules run in code on every string that ships in a row (user and assistant turns, reasoning content, and every other field), and system prompts are checked separately.

Hard rules (any hit removes the row): email, phone number (US and international), street address, SSN shape, Luhn card number, a labelled ID (passport, licence, date of birth and similar), a name with an age, "lives with", a family detail, the pipeline's own PERSONAL_DETAIL pattern, the short list of private people from the Stage B build, and these rules added after the red team review:

• family relations: a name followed by a possessive family word ("X's son", "X's wife");
• birth years with names: "born 1984" style dates in a row that names a person;
• unnamed minor suspects: a minor (an age under 18, teen, juvenile, minor, boy, girl, youth, pupil) in the same sentence as a suspect or justice cue (arrested, charged, accused, detained, sentenced, shooter, attacker and similar);
• named killings: "killing of", "murder of", "death of", "shooting of" and similar followed by a name, anywhere in the row;
• incident articles: every row from a Wikipedia article whose title names a violent incident or disaster (shooting, stabbing, massacre, murder, killing, kidnapping, bombing, attack, arson, fire, crash, collapse, stampede, explosion, disappearance, assassination, riot, hostage, ramming), and every row from a section about victims, perpetrators, suspects, investigations, trials, arrests, casualties, deaths, personal life, early life or family on any article;
• named killing and case articles: titles such as "Killing of ...", "Murder of ...", "Trial of ...", or "X v. Y".

Soft rule (kept on purpose, it over-drops): a personal name in the same sentence as a crime, minor or resident cue. About 1 in 10 of its hits named a private person in the review sample, so it removes many rows about public figures too.

Hard-rule hits by rule, over all input rows (a row can hit several): family_detail 1,601, id_labelled 1,109, age_with_name 949, named_killing_text 865, street_address 711, family_relation 694, birth_year_with_name 399, unnamed_minor_suspect 290, PERSONAL_DETAIL 99, email 38, phone 35, lives_with 14, card_luhn 3, ssn_shape 1.

Named people in a justice context (added by the second read): a personal name together with an execution, death sentence, death row, detention, imprisonment or forced-confession cue, or with an under-18 offender cue, in the same sentence (or anywhere in rows under 1,200 characters). This over-drops on purpose and also removes rows about public figures.

Court records are removed entirely, whatever else the row holds: the s3-courtlistener and us-scotus files, and, after the second read, any row whose source record is a court judgment, court order, Advocate General opinion or preliminary-ruling request (the EUR-Lex Court of Justice documents).

Residual risk: before these rules were widened, a manual read of 320 random rows found 2 naming a private individual (0.62%, 95% Wilson interval 0.17% to 2.25%) and 10 naming non-public people in a work role (spokespeople, a school principal, analysts). The widened rules remove the categories the red team found, but regex rules miss some names, so a small number of private individuals may remain. Report any you find (see "Takedown") and the rows will be removed. A second, fresh sample of 441 rows from the widened output (none from earlier samples; 26 reasoning rows checked by metadata only) found 2 of 415 rows read in full naming a private individual (0.13% to 1.74%, 95% Wilson interval). Both rows, and the class they belong to (rule above), were removed. Rows written in other languages are covered less well, since several rules assume English.

Verbatim guard

Training rows never contain source passages: the user turn is the question alone. Answers and questions can still reuse wording from the source. For every row whose source is publisher text and whose source record is on the build machine, the build compared each shipped string (question, answer, reasoning content and other fields) with the record's title and text, and removed the row if any run of 12 or more consecutive words matched (case-insensitive, word tokens). The second read ran the same 12-word guard on every noncommercial row (NC sources are not open text either), and a 40-word passage guard on every row from an open-license collected source, so no long source passage ships under any license. Wikipedia rows were not checked: the article text is not on the build machine. Short quotes from open sources (under 40 words) remain, under those sources' licenses.

• Checked: 568,133 rows. Removed: 193,889 (34.1%).
• Removed by input file: stage2 162,655, stage1 22,855, stage_b_r30b 7,347, stage_b_written 795, measured 205, eval_honesty_dev_120 29, eval_probes_2500 2, eval_probes_1500 1.
• Removed by source: s3-ccnews-general.jsonl 100,119, s3-ccnews-ai-meta.jsonl 57,853, pytorch-blog.jsonl 8,347, s3-ccnews.jsonl 7,556, meta-docs.jsonl 4,737, meta-engineering.jsonl 3,517, meta-newsroom.jsonl 3,273, us-sec-8k.jsonl 2,571, us-sec-meta.jsonl 1,554, s3-ccnews-plus.jsonl 1,096, s3-openalex.jsonl 597, github-meta-orgs.jsonl 547, anthropic-release-notes.jsonl 385, gap-meta-dev-blog.jsonl 377, un-news.jsonl 255, gap-hf-facebook.jsonl 234, gdacs.jsonl 230, agencia-brasil.jsonl 215, s3-forecast-news.jsonl 188, global-voices.jsonl 157, meta-cards.jsonl 66, s3-pmc.jsonl 11, launch-library.jsonl 2, medrxiv.jsonl 2.
• Not checked at build time: stage1 3,172 (anchor), stage1 3,687 (missing record), stage2 3,172 (anchor), anchors 3,172 (anchor).
• Missing records: stage 1 CC-NEWS rows (s3-ccnews-ai-meta) whose source records were not kept on the build machine, so the guard could not run on them. Of the 7,891 stage 1 rows from the same source that it could check (a build-time count; the removed rows are not in this release), 3,850 are published after the guard and the second read. All 3,686 missing-record rows still in the build after the second read were removed (the second read had removed 1). No published row skipped the guard for a missing record.
• Anchors were written from Q&A pairs, not from source text, so there is no single source to compare against. They are published marked unchecked_anchor: spark/anchors 3,128, spark/stage1 3,128, spark/stage2 3,128.

The stage-b honesty dev set's facts field (up to 3,000 characters of source passage) is removed, and so are source URLs, except the source_url kept on evals/salient_246 rows (see "The release object on every row").

Second read

After the build, a second reader drew a fresh random sample across every split, read it by hand, grepped every file in full, and removed rows with these rules (counts are rows removed, by first rule that hit):

• source record is a court judgment, order or opinion: 1,368
• open-source row shares 40+ consecutive words with its source record: 687
• noncommercial row shares 12+ consecutive words with its source record: 649
• named person in a justice, detention or execution context: 543
• grep: street address with zip code (reasoning row, classified in code): 2
• hand check: private individual (relationship detail): 1
• hand check: private individual (detention context): 1

Sample results (95% Wilson intervals, in percent of rows):

Check	Hits	Rows	95% interval
private individual personal data	2	415	0.13 to 1.74
non public person in work role	3	415	0.25 to 2.1
raw source text 12w from non open source	10	415	1.31 to 4.38
raw source text 40w any source	0	415	0 to 0.92
secrets	0	441	0 to 0.86
exa content	0	441	0 to 0.86
wrong license tag	0	441	0 to 0.86
court record	0	441	0 to 0.86
nd or no reuse	0	441	0 to 0.86

Full-file grep of every split: no email addresses, phone numbers, AWS keys, account ids, hf_ tokens, sk- keys or presigned URL fragments; IP addresses are loopback, private-range or version numbers; file paths are ones quoted from public docs, none from the build machine; no Exa or LittleBit content.

Who wrote the rows

• Muse Glimmer 30B, base model, self-hosted with vLLM (RedHatAI/Muse-Glimmer-30B-INT4 revision 397b1ee; FP8 build 1deb464), temperature 0.4 for facts and 0.7 for Stage B items. Apache 2.0.
• Muse Spark Open on Fireworks, Baseten and Modal: most stage 1 fact rows, the anchors, the 1,500 fixed probes and 25 named controls.
• stage2-multi (the tuned stage 2 model): the measured answers.
• Templates written in code: cutoff and after-cutoff replies, measured-wrong replies.
• The project repo: the honesty 190, salient 246, reasoning 260, old facts 40 and cutoff 20 sets.
• No Meta Model API output, Exa result or LittleBit output is in this release.

Evaluation sets and contamination

Every eval and test set behind stage-b-r30b is in evals (one NC-sourced honesty dev row is in noncommercial, still tagged as an eval). All of them are contaminated for any model trained after this release: future runs need new test sets, and scores on these sets compare only models trained before it. Each eval row says so in release.eval.

Split	What it is	Published
probes_1500	Fixed short-answer probes, one per passage, never trained on as questions (Spark Open)	1,433
probes_2500	The 1,500 plus 1,000 from passages stage 1 never trained on	2,426
honesty_dev_120	Honesty dev: false premise, after cutoff, unknowable (base Muse Glimmer)	82
named_controls_150	Real named things Stage B never trains on, to measure over-refusal	148
honesty_190	Held-out honesty test (70 after cutoff, 90 false premise, 30 unknowable), written in the repo	189
salient_246	Salient 2026 events from outside lists (Wikipedia year pages, Meta newsroom)	243
reasoning_260	Multi-step and skills items generated from a seed	260
old_facts_40	Facts from before 2026, to check forgetting	40
cutoff_20	Cutoff questions	20

Known limits of these sets:

• Every one of the 2,500 probes comes from a passage the model trained on. They measure reworded questions about trained passages, not coverage of 2026.
• The honesty dev set came from the same writer as the Stage B rows and overstates honesty (85.8% honest) against the held-out 190 (44.7%, 95% CI 38 to 52, blind hand audit).
• One old-facts question also appears word for word in a training file.
• The measured questions (model-outputs/measured) were held out of training as questions, but their answers were used to pick Stage B rows, so treat them as contaminated too.

Limitations

• Machine-written: questions and answers can be wrong. Checks are code-only, and stories that changed later can have outdated answers.
• The removals are not neutral: the verbatim guard removes about a third of checked news rows, and the personal-data rules remove more crime, disaster and conflict coverage than other topics.
• Dates: the trained cutoff templates say October 28, 2026, while Stage B restates October 29, 2026.
• On the held-out honesty set the tuned model still accepts 80% of false premises.
• Mostly English.

Takedown

If a row contains personal data about you, text you hold rights in, or anything else that should not be here, open a discussion in this dataset's Community tab on Hugging Face, giving the row's release.from_file and release.line. You don't need to quote the row. Rows are removed from the next revision of the dataset.