docs/evals/m1_data_eda.md

M1 data EDA — UniversalCEFR

Generated: 2026-06-11T14:28:37+00:00 Command: scripts/eda_universalcefr.py

Subsets overview

dataset	lang	rows	categories	formats	licenses	words: median (p10, p90)
UniversalCEFR/cambridge_exams_en	en	331	{'reference': 331}	{'document-level': 331}	{'CC BY-NC-SA 4.0': 331}	563 (p10=123, p90=923)
UniversalCEFR/cefr_asag_en	en	299	{'learner': 299}	{'document-level': 299}	{'CC BY-NC-SA 4.0': 299}	59 (p10=8, p90=150)
UniversalCEFR/cefr_sp_en	en	10004	{'reference': 10004}	{'sentence-level': 10004}	{'CC BY-NC-SA 4.0': 10004}	14 (p10=8, p90=24)
UniversalCEFR/elg_cefr_en	en	712	{'reference': 712}	{'document-level': 712}	{'CC BY-NC-SA 4.0': 712}	277 (p10=191, p90=799)
UniversalCEFR/icle500_en	en	495	{'learner': 495}	{'document-level': 495}	{'CC0 1.0 (Public Domain)': 495}	585 (p10=476, p90=762)
UniversalCEFR/readme_en	en	2822	{'reference': 2822}	{'sentence-level': 2822}	{'CC BY SA NC 4.0': 2822}	17 (p10=7, p90=36)

Level distribution per subset

dataset	A1	A2	B1	B2	C1	C2	odd labels
UniversalCEFR/cambridge_exams_en	0	64	60	71	67	69	0
UniversalCEFR/cefr_asag_en	18	59	113	74	30	5	0
UniversalCEFR/cefr_sp_en	124	1271	3305	3330	1744	230	0
UniversalCEFR/elg_cefr_en	24	110	158	140	97	69	114
UniversalCEFR/icle500_en	0	1	29	99	91	64	211
UniversalCEFR/readme_en	182	673	623	896	377	71	0

Odd labels detail: {'UniversalCEFR/elg_cefr_en': {'A1+': 1, 'A2+': 31, 'B1+': 48, 'B2+': 34}, 'UniversalCEFR/icle500_en': {'B2+': 119, 'B1+': 85, 'NA': 7}}

Reference pool (the M1 reading-classifier candidates)

Subsets whose category includes reference, i.e. texts written for learners rather than by them. Learner-production subsets are the M3 candidates (grading learner writing), not M1 training data.

lang	reference rows	from subsets
en	13869	cambridge_exams_en, cefr_sp_en, elg_cefr_en, readme_en

Pooled level distribution for en reference subsets (class balance check):

A1	A2	B1	B2	C1	C2	odd
330	2118	4146	4437	2285	439	114

---

Notes: word counts use whitespace tokenisation (approximate for ar/hi). Licenses are aggregated from the per-row license field; decisions and exclusions are recorded in docs/adr/0003-datasets-and-licensing.md.