EN Translation — model comparison

Five models translated the same PT→EN content. Here the same excerpt comes through each one, grouped by linguistic domain (each domain stresses a different skill). Adopted: Sonnet (EN4) — the site selector is just PT · EN. The live variants have been retired; the examples below remain as a record.

ModelType
EN1qwen2.5-coder:14blocal (Ollama)
EN2qwen3-coder:30blocal (Ollama)
EN3Claude Haiku 4.5cloud
EN4Claude Sonnet 4.6cloud
EN5Claude Opus 4.8cloud

1. Personal prose / narrative

Tests register and voice — the intimate tone must not sound robotic.

PT: Luto diariamente para superar a minha inevitável temporalidade deixando coisas boas por onde passo…

  • EN1: I grieve daily to overcome my inevitable mortality, leaving good things behind wherever I go… → ❌ “grieve” (luto = to struggle, not to grieve) and “mortality” (≠ temporality): two meaning errors.
  • EN2: I grieve daily to overcome my inevitable temporality, leaving good things in my wake… → “temporality” correct, but keeps “grieve”.
  • EN3: I fight daily to overcome my inevitable temporality… → correct and natural.
  • EN4: I struggle daily to overcome my inevitable temporality by leaving good things wherever I go… → ideal register.
  • EN5: I struggle daily… → identical to EN4 in quality.

Verdict: EN4 ≈ EN5 > EN3 > EN2 > EN1 (meaning error).


2. Philosophical / abstract

Tests conceptual precision and fidelity to an argument (Plato, discovery×creation).

PT: A fórmula já existe; nós descobrimos (Platão). A intenção preserva a descoberta crua; as saídas são a materialização. A regra emerge do uso — de baixo pra cima — não imposta.

  • EN1: …we discovered (Plato). The intent… the manifestation… from bottom to top… → drops the “it”, intent (≠ intention), manifestation weak.
  • EN2: …we discovered it (Plato)… its materialization… from bottom to top… → correct; “bottom to top” stilted.
  • EN3: …we discover it (Plato)… bottom-up… not imposed. → clean and idiomatic.
  • EN4: …we discover it… bottom-up — never imposed. → more emphatic and faithful (“não imposta”).
  • EN5: …we discover it… from the bottom up — not imposed. → excellent, slightly more formal.

Verdict: EN4 ≈ EN5 > EN3 > EN2 > EN1.


3. Instructional / technical

Tests terminology (acronyms: ODS→SDG) and clarity of instruction.

PT: Iniciar a exploração dos Objetivos de Desenvolvimento Sustentável da ONU (vídeo de introdução, vídeos por ODS de interesse, e conclusão)

  • EN1: …videos on ODS of interest… → ❌ left the acronym ODS untranslated (should be SDG).
  • EN2: …videos of interest by SDG… → correct; word order slightly awkward.
  • EN3: …videos per SDG of interest… → clear and correct.
  • EN4: …videos per SDG of interest, and wrap-up → clear; “wrap-up” too casual for “conclusão”.
  • EN5: …videos for SDGs of interest, and conclusion → more precise.

Verdict: EN5 > EN3 ≈ EN2 > EN4 (register) > EN1 (acronym).


4. Conversational / dated

Tests style and conventions — the ×, the intentional lowercase “yuri”, no Title-Case.

PT: Reunião yuri × Miguel: conteúdo × forma, método de estudo e experiências de aprendizado (login/autenticação), com warm-up sobre jogos para a próxima semana.

  • EN1: Yuri and Miguel Meeting: Content × Form, Study Method… → ❌ Title-Case everything and lost the × in the name — clashes with the style.
  • EN2: Yuri × Miguel meeting: content × form… → good; capitalized “Yuri”.
  • EN3: Meeting yuri × Miguel: content × form… → preserves the lowercase.
  • EN4: yuri × Miguel meeting: content × form… → preserves style and natural order.
  • EN5: yuri × Miguel meeting: content × form, study method**,** and… → same, with Oxford comma.

Verdict: EN4 ≈ EN5 ≈ EN3 > EN2 > EN1 (style).


5. Domain terms / proper nouns

In all: names (Alpha Scholars, Minecraft, Stanford, MS&E) and the tag glossary (board, garden, missions) came out identically — this is handled by fixed rules, not by the model. EN1 also tends toward Title-Case on short titles.


Recommendation (opinionated)

Adopt EN4 — Claude Sonnet 4.6. This universe has a reflective and philosophical voice; what matters is register and conceptual fidelity, not just literal meaning. Sonnet nails the tone (“struggle”, not “grieve”), preserves style choices (lowercase “yuri”, no Title-Case), is idiomatic and faithful (“never imposed”), and is faster/cheaper than Opus with only marginal difference.

  • Premium: EN5 (Opus 4.8) — ties with Sonnet, slightly more formal. Use if you want the absolute quality ceiling.
  • Offline: EN2 (qwen3-coder:30b) — best local option, no cloud; minor rough edges (“grieve”, “bottom to top”).
  • Avoid: EN1 (qwen2.5-coder:14b, the current one) — has real meaning errors (grieve/mortality), untranslated acronym (ODS) and unwarranted Title-Case.

One-line decision: promote EN4 (Sonnet) to canonical /en/; keep the rest as comparable form here.