Hy-MT2-1.8B Discord (LoRA)

LoRA adapter (r=32) for tencent/Hy-MT2-1.8B (Apache-2.0) that translates Discord messages while keeping markdown, links, mentions, emojis, code and [markNN] tags (variable number) intact. Merged weights and a bf16 GGUF are in Jour/hy-mt2-1.8b-discord. A LoRA for the official 1.25-bit GGUF is in Jour/hy-mt2-1.8b-discord-1.25bit-lora.

Modified from the original: weights fine-tuned on Discord-style translation pairs. The Apache-2.0 license of the base model is included (LICENSE.txt).

Results (143 unseen pairs, 108 containing protected elements)

formatting preserved chrF
base bf16 94.4 % 63.5
+ this LoRA 97.2 % 70.6

"Formatting preserved" = every protected span (tags, links, mentions, emojis, inline/fenced code) is identical in the output, markdown markers and line breaks counts match.

Usage

Use the official Hy-MT2 prompt, no system prompt: Translate the following text into French. Note that you should only output the translated result without any additional explanation:\n\n{text} (Chinese prompt for zh).

Training data & limits

~2.9k synthetic pairs generated by an LLM (Claude Haiku) and automatically validated (identical protected spans between source and translation), 14 languages, heavily skewed towards English and French. No real user messages. Translations are not human-reviewed; small test set; keep a formatting check (tag masking) in your application.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Jour/hy-mt2-1.8b-discord-lora

Adapter
(4)
this model