HarleyCooper/riggs1890-lexicon-graph

Riggs 1890 Lexicon Graph

One dictionary volume as a provenance-linked graph, with the supervised rows and RL tasks that are built from it.

• Source: Stephen R. Riggs, A Dakota-English Dictionary (1890). The book is in the public domain.
• Code: https://github.com/HarleyCoops/Dakota1890 (lexgym/). Two commands rebuild every file here from the per-page extraction.
• Environment: environments/riggs1890_lexgym in the same repository (a verifiers package).
• Traces: HarleyCooper/riggs1890-lexgym-traces
• Pilot adapter: HarleyCooper/Qwen3.5-4B-Riggs1890-lexgym-pilot

Configs

Config	Rows	Content
dakota_entries	16,980	Dakota→English half (scan pages 95–370). Headword with syllable division and accent, plain form, part of speech, gloss, inflected forms, derivation links, word family, split.
english_entries	7,244	English→Dakota half (scan pages 371–430). English headword, Dakota equivalents, and which of them are also headwords in the other half.
edges	12,931	derived_from, see_also, and english_equivalent links, with a flag for links that resolve to an entry.
rl_tasks	68,911 train, 5,284 held-out, 600 recall	Eight task types with a gold answer, an accepted set, and provenance.
sft	56,120	Chat rows for the supervised warm start (six single-turn task types).

lookup_index.json holds the text that the dictionary lookup tool returns for each headword.

How it was built

A vision model read 336 scan pages into 24,224 structured entries. The graph builder then applies four rules.

1. Canonical orthography. The vision model wrote one printed letter with several Unicode characters. A fixed table maps the variants to one form. headword_raw keeps the form as it was read. A form with a character outside the Riggs alphabet is kept but is never a gold answer.
2. Trusted headwords. headword_ok is true when the printed syllable division and the single accent mark parse without error (16,227 of 16,980).
3. Two witnesses. A Dakota form in the English half that is also a headword in the Dakota half has two independent readings (5,279 of 10,779 forms). English→Dakota RL and evaluation tasks accept only these forms.
4. Word families. A headword and all words derived from it are one family (12,135 families). One family in ten is held out. No held-out family is in a training row.

Task types

Task	Input → answer	Reward
syllabify	plain word → Riggs' syllable division and accent	exact; partial credit for boundaries and stress; 0 if a letter changes
pos	word → part-of-speech abbreviation	exact
inflect	word → the inflected forms Riggs lists	set F1
root	derived word → the word it is derived from	set F1
gloss	Dakota word → English definition	content-word F1
en_to_dak	English word → Dakota word	member of the accepted set
lookup_gloss	multi-turn, one dictionary read needed	content-word F1 on the final answer
lookup_root_gloss	multi-turn, two dictionary reads needed	content-word F1 on the final answer

Splits: heldout_rule uses word families that training does not contain, and measures generalization. recall uses training vocabulary, and measures memory of the volume.

Every row has entry_id, page, and source_image, so each fact can be checked against the scan.

Mixed SFT file (config sft_mixed)

tasks_jsonl/sft_mixed.jsonl has 69,461 rows: the 56,120 single-turn rows of sft plus 13,341 rows from solved lookup conversations (3,000 one-hop and 2,447 two-hop). tasks_jsonl/build_mixed_sft.py builds it from rl_train.jsonl and lookup_index.json with no teacher model. Each conversation is replayed against the environment rules and kept only if it scores 1.0. It is written as one row for each assistant turn, so the loss is on the last assistant message of a row.

Training on this file kept tool use. thinkingmachines/Inkling-2-Small, three passes, exact match on 1,500 evaluation tasks:

Measure	Base	Pass 3
Recall from memory	0.030	0.835
Held-out word families	0.443	0.816
Lookup, one hop	0.767	1.000
Lookup, two hops	0.893	1.000

Method: cookbook. Live demo: HarleyCooper/riggs1890-lexgym. Traces: riggs1890-lexgym-traces.

Use

from datasets import load_dataset

tasks = load_dataset("HarleyCooper/riggs1890-lexicon-graph", "rl_tasks")
entries = load_dataset("HarleyCooper/riggs1890-lexicon-graph", "dakota_entries", split="train")

Limits

• The source is a missionary dictionary from 1890. It records one written variety and is not a description of Dakota as it is spoken now.
• The extraction has errors. 1,494 forms in the English half differ from a headword only by n / ŋ, which is likely a reading error. 753 headwords failed the parser.
• The orthography mapping is a set of hypotheses about the extraction. It is listed in lexgym/orthography.py.
• A model trained on this data is a starting point for correction by Dakota speakers. It does not replace them, and this dataset carries no community endorsement.

License

The 1890 text is in the public domain. The structured files and the code are released under Apache-2.0, the license of the repository.