BowLingual-E4B-v1.5

BowLingual-E4B-v1.5 is a fine-tuned Gemma 4 E4B model specialized for English → natural conversational Thai translation, particularly informal messaging such as LINE and Messenger.

The goal is native-sounding Thai chat rather than literal or textbook-style machine translation.

Training data

The final SFT dataset contained 17,437 exact-deduplicated examples.

BowLingual curated gold

• 4,932 examples
• Original curated BowLingual English/Thai training data

Wisesight-derived

• 8,095 examples
• Original human-written Thai from the Wisesight sentiment corpus was preserved
• English source messages were reconstructed from the Thai using Gemma 4 31B
• Candidate reconstructions were independently scored and selected
• Thai text was never rewritten by the teacher model

SCB-MT-derived

• 4,410 examples after exact deduplication
• Original English and Thai from SCB-MT-EN-TH-2020 were preserved
• Rows were selected using strict Thai-naturalness, semantic-accuracy, role-accuracy, modernness, and chat-suitability judging

Training

• Base: unsloth/gemma-4-E4B-it
• Precision: BF16 LoRA
• LoRA rank: 16
• LoRA alpha: 16
• Trainable parameters: ~36.7M
• Sequence length: 2048
• Effective batch size: 16
• Epochs: 3
• Initial learning rate: 1e-4
• Scheduler: cosine
• Warmup: 3%
• Loss calculated on assistant responses only
• Best checkpoint selected by validation loss

Intended use

BowLingual-E4B-v1.5 is intended primarily for translating casual English messages into natural conversational Thai.

It is not optimized for legal, medical, technical, or other formal translation.

Dataset provenance

Relevant upstream dataset licenses and terms still apply to their respective source material. In particular, SCB-MT-EN-TH-2020 is distributed under CC BY-SA 4.0. Review upstream dataset terms before redistribution or commercial use.