opencode-translate 0.3.6 → 1.0.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +4 -1
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -2,13 +2,16 @@
|
|
|
2
2
|
|
|
3
3
|
|
|
4
4
|
|
|
5
|
+
## Demo
|
|
6
|
+
|
|
7
|
+
|
|
8
|
+
|
|
5
9
|
## Why
|
|
6
10
|
|
|
7
11
|
LLMs are worse in non-English. Benchmarks confirm it on every frontier model.
|
|
8
12
|
|
|
9
13
|
- **Claude Opus 4.7 / GPT-5.5 / Gemini 3.1 Pro**: identical coding tasks scored **5/5 in English** but dropped to **0–1/5 in Arabic and Korean** ([LILT, 2025](https://lilt.com/blog/multilingual-ai-coding-gap-non-english-developers)).
|
|
10
14
|
- **Anthropic's own numbers**: Japanese 96.9%, Korean 96.6%, Yoruba 80.3% — English is always the baseline ([Anthropic docs](https://platform.claude.com/docs/en/build-with-claude/multilingual-support)).
|
|
11
|
-
- **Translation Barrier Hypothesis** (JHU, IJCNLP 2025): LLMs reason in an English-dominant latent space, then translate at the last layers — every translation step is a source of error.
|
|
12
15
|
- **Token tax**: Korean ~1.25×, Japanese ~1.25×, Arabic ~3× more tokens per equivalent content. Higher cost, smaller effective context.
|
|
13
16
|
|
|
14
17
|
This plugin lets you write in your language while the model works in English — best of both worlds.
|
package/package.json
CHANGED