Under the fixed protocol, hosted LLMs produce the largest gain on Spanish thematic classification, with thematic macro F1 rising from 0.46 (TF-IDF) to roughly 0.82-0.86 for the two hosted LLMs¶
Claim · All claims
Evidence · 1 study · designr2·q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i? On the 233-comment Spanish held-out set, hosted LLM arms reach thematic macro F1 of roughly 0.82-0.86 versus 0.46 for TF-IDF and about 0.68 for frozen encoders. → Esteban U. Vega Barajas 2026
Evidence¶
Esteban U. Vega Barajas 2026¶
Esteban U. Vega Barajas. (2026). A Durability and Cross-Language Transfer Benchmark for a Validated Teaching-Feedback Classification Protocol. https://arxiv.org/abs/2607.11873 doi:10.48550/arxiv.2607.11873
q2 · i? · design · r2
Descriptive comparison across arms on the unified 233-comment Spanish held-out split (seed 20260616), all arms sharing the identical labeled subset, guide, and held-out set. The article reports the thematic macro F1 jump "from 0.46 for TF-IDF to about 0.68 for BETO and the encoder, then to roughly 0.82 to 0.86"; no significance is claimed.
"The largest single jump is in thematic macro F1: it rises from 0.46 for TF-IDF to about 0.68 for BETO and the encoder, then to roughly 0.82 to 0.86 for the two hosted LLMs."
Discussion¶
Related Claims¶
- Sentiment classification transfers to English under the unchanged protocol, with a frozen multilingual encoder the strongest arm at wF1 0.73 and LLMs matching it only with few-shot prompting — related
- The frozen BETO-2019 design and the modern e5-base encoder are numerically close on Spanish, with no descriptive separation on either task — related
- The small open-weights Qwen2.5-1.5B model is the weakest arm in both languages, with Spanish thematic weighted F1 of 0.31 zero-shot rising to 0.63 with few-shot — related
- On Spanish sentiment, the 2026 frontier model shows no advantage over the cheap model, with Haiku zero-shot at 0.923 versus Opus zero-shot at 0.884 and the head-to-head McNemar favoring Haiku — related
- Self-hosted open-weight LLMs offer privacy, reproducibility, and environmental advantages over commercial LLM interfaces in educational ITSs — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
llms-largest-gain-spanish-thematic-macro-f1 |
title |
Under the fixed protocol, hosted LLMs produce the largest gain on Spanish thematic classification, with thematic macro F1 rising from 0.46 (TF-IDF) to roughly 0.82-0.86 for the two hosted LLMs |
description |
Under the fixed protocol, hosted LLMs produce the largest gain on Spanish thematic classification, with thematic macro F1 rising from 0.46 (TF-IDF) to roughly 0.82-0.86 for the two hosted LLMs |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — esteban-u-vega-barajas-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.