slm-lipogram

Verified training examples for teaching a small model to rewrite English text without the letter "e". The default config, us_open1 (150 rows), trained the final model jwjesse/no-e (Qwen3-1.7B, QLoRA). v2 and v1 (188 BBC news rows each) are the previous version, which trained jwjesse/slm-lipogram-v2e10-n188. Every rewrite passed a code check for no "e" outside names copied exactly from the source, real English words, length and kept names, and a meaning judge from a different provider than the teacher. Project and results: https://labs.gauntletai.com/jessewells/SLM (by Jesse Wells; contact via the GitLab repo).

Passage-to-rewrite pairs for training a small model to rewrite English text without the letter "e" (either case) while keeping every key fact. Each row is a chat example: the user message is Rewrite this passage without using the letter e: followed by a news passage, and the assistant message is an e-free rewrite.

{"id": "lip-...", "source_id": "...", "messages": [{"role": "user", "content": "Rewrite this passage without using the letter e:\n\n<passage>"},
 {"role": "assistant", "content": "<rewrite>"}], "meta": {"repair_rounds": 0, "judge_score": 5, "e_count_first_draft": 0}}

us_open1 rows also carry meta.source (cnn or wikipedia) and meta.holistic_meaning (1 to 5):

{"id": "lip-usopen-wikipedia-4514132", "source_id": "usopen-wikipedia-4514132", "messages": [...],
 "meta": {"repair_rounds": 0, "judge_score": 4, "holistic_meaning": 5, "e_count_first_draft": 0, "source": "wikipedia"}}

"Stray e" in the project’s results means an "e" outside names copied exactly from the source passage; the rule check masks such names before counting.

Versions and how to load

Config	Rows	Notes
us_open1 (default)	150	Openings of US CNN articles and Wikipedia introductions, written by gpt-6.1-sol. 75 CNN, 75 Wikipedia, seed-0 shuffle, so the first N rows are a valid size-N subset (N = 19, 38, 75). See the next section.
v2	188	128 v1 rows plus 60 name examples (passages whose named entities contain an "e"; see below). Seed-0 shuffle, so the first N rows are a valid size-N subset.
v1	188	Passages with no named entity containing an "e" (filter below). Seed-0 shuffle, so the first N rows are a valid size-N subset (N = 25, 50, 100).

from datasets import load_dataset
ds = load_dataset("jwjesse/slm-lipogram")                 # us_open1, the default
ds_v2 = load_dataset("jwjesse/slm-lipogram", "v2")        # previous version
ds_v1 = load_dataset("jwjesse/slm-lipogram", "v1")        # v1

us_open1 (default): US openings

150 examples. The user prompt is the same as above, but the passages are US English: the opening of a CNN article or a Wikipedia introduction, 50 to 100 words cut at a sentence end. Each row has meta.source (cnn or wikipedia).

• Sources and licences: CNN articles from abisee/cnn_dailymail 3.0.0 (only articles naming "(CNN)" up front; Daily Mail excluded; the dataset is listed as apache-2.0, the article text belongs to CNN) and English Wikipedia from wikimedia/wikipedia 20231101.en (CC BY-SA 3.0 and GFDL; share-alike, so passages and rewrites here need the same terms and an attribution to the Wikipedia authors). Check the rights position before any use beyond research. The rewrites were generated by a model.
• Openings filter: the CNN lead after the "(CNN) --" dateline, or the Wikipedia introduction. Dropped: unbalanced quotes, a first word that refers back (He, It, But...), markup, URLs, lists, inline image captions, odd spacing, words valid only in British English, or a low English share. Only train-split documents are used; none of the 150 eval passages is in this data.
• Haiku screen: CNN text keeps photo captions inline, so Claude Haiku 5.5 dropped passages with captions, sentences restating the lead, bio or sidebar facts, or sentences that do not follow from the previous one.
• Teacher: OpenAI gpt-6.1-sol through OpenRouter, reasoning effort high, name placeholders as in v2, up to 3 repair rounds.
• Gates: no "e" outside names copied exactly from the source (86 of 293 attempts rejected), ASCII letters only, real English words, length at least 70% of the source, meaning judge (Claude Sonnet 5.5) score at least 4 of 5 (51 rejected), then a holistic meaning check (Sonnet 5.5, score 4 or more; none rejected). 156 of 293 attempts accepted (53.2%); 150 used.
• Cost: about $11.31 for the teacher and $1.51 for the judges.

Older versions (v2, v1)

The sections below describe v1 and v2, built from BBC news (XSum). They are kept as the previous version.

How it was made (v1 and v2)

• Source passages: train-split passages (50-100 and 100-150 words, runs of consecutive lines) from the document field of EdinburghNLP/xsum, BBC news articles.
• v1 filter: a passage is skipped if it has a capitalised word containing "e" (a name, month, weekday) or an all-caps token with "e", other than a common word that only starts a sentence. This keeps names out of the hardest part of the task: 350 of 6,000 candidate passages survive.
• Teacher: OpenAI gpt-6.1-terra, reasoning effort high, with a lipogram system prompt. A draft that still has problems goes back with the offending words for up to 3 repair rounds.
• Gates, in order (all must pass): no "e" outside names copied exactly from the source (look-alike letters such as Cyrillic "e" count; 5 rejected), ASCII letters only (1 rejected), only real English words or words from the source (1 rejected), length at least 70% of the source word count (0 rejected), meaning judge score at least 4 of 5 (74 rejected). 188 accepted (69.9%). Counts above are v1 (269 attempted). For the v2 name rows (140 attempted): name placeholders kept 140 of 140, no "e" outside names 133 (7 rejected), ASCII 130 (3), rewrites with only real words 129 (1), length 129 (0), meaning 64 (65 rejected). 64 accepted (45.7%), of which the first 60 are used. Mean judge score of accepted rows: 4.08.
• Language filter: XSum includes Welsh articles, so a passage needs at least 40% of its words in the en_US dictionary. No row in v1 or v2 is affected (lowest share 0.83); 32 of 9,000 candidate passages were skipped.
• Judge: Claude Sonnet 5.5 (claude-sonnet-5-5, thinking off): 1 to 5 on whether all key facts survive. Mean score of accepted rows: 4.11. The pipeline and the rule check are in the project repository (lipogram/generate.py, spec/).
• Names without an "e" in a source passage may be kept as written; names with an "e" do not occur in the v1 passages.

Licences at a glance

The YAML says license: other because the sources differ. CNN text belongs to CNN (via abisee/cnn_dailymail, itself listed as apache-2.0). Wikipedia text is CC BY-SA 3.0 and GFDL: attribute the Wikipedia authors and share alike. XSum/BBC text (v1, v2) has an unknown licence; research use only. The rewrites were generated by a model. For us_open1, source_id is usopen-<source>-<id>: for Wikipedia the id is the wikimedia/wikipedia article id (https://en.wikipedia.org/?curid=<id> opens it); for CNN it is the abisee/cnn_dailymail article hash. The project repo’s lipogram/data/us_open_pool.jsonl maps each to its article_id, doc_id and licence.

Source and licence (v1 and v2)

The source text is from XSum (EdinburghNLP/xsum), BBC news articles. The licence is listed as unknown on the XSum dataset card, and the text belongs to the BBC. The passages are included here for research use; check the rights position before any other use. The rewrites were generated by a model.

Citation for the source dataset:

@inproceedings{narayan-etal-2018-dont,
  title = "Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization",
  author = "Narayan, Shashi and Cohen, Shay B. and Lapata, Mirella",
  booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
  year = "2018", publisher = "Association for Computational Linguistics"
}

v2: examples with names

v2 keeps the size of v1 (188 rows). It drops 60 v1 rows, chosen with random.Random(0).sample(range(188), 60) (the dropped ids are listed in v2_dropped_ids.json in the project repository), and adds 60 rows made from passages that v1 skipped because they contain named entities with an "e".

• Rule for names: an "e" is allowed only inside a named entity copied exactly from the source (spaCy en_core_web_sm: PERSON, GPE, LOC, ORG, NORP, FAC, EVENT, plus day and month names). Whole entities count; a person’s name may be shortened to a part such as the surname; other entities only to their last capitalised word. Generic words and titles (Party, Royal, Queen, Parliament...) never count alone.
• How the names are kept exact: before the teacher sees a passage, each such name is replaced by a placeholder ([N1], [N2]...) that the teacher must keep once and unchanged, and the names are put back before any check. A draft with a missing, duplicated or altered placeholder is repaired, then rejected. No placeholder appears in the data.
• Same teacher, judge and other gates as v1; the gate counts for the 60 name rows are listed under How it was made.