GLOSS-MT: Grounded Lexical Offline Search for Sequence Translation (Gemma 4 E4B)

GLOSS-MT (ghanaopenai/gloss-mt) is a 4-billion parameter state-of-the-art neural machine translation model for English to Asante Twi (Akan), fine-tuned on google/gemma-4-E4B-it. It pioneeringly combines autonomous dictionary tool-calling with explicit linguistic reasoning to solve lexical scarcity and hallucination in low-resource machine translation.

• Author: Ghana Open AI
• Supported by: Ghana NLP
• GitHub Repository: GhanaOpenAI/gloss-mt
• Interactive Live Demo: Hugging Face Spaces: twi-translator
• Base Model: google/gemma-4-E4B-it (4B parameter compact multimodal/language foundation model)
• Bundled Dictionary: 545,633 English→Twi verified offline entries

---

Benchmark Evaluation (nsanku-mt-benchmark)

Evaluated across the comprehensive document-level nsanku-mt-benchmark suite (84 documents spanning 4 heterogeneous domains) against leading commercial and open-source translation models:

Overall Leaderboard (English → Asante Twi)

Model	Track / Architecture	Parameters	chrF2 (Avg)	BLEU	Finance (3 Docs)	UDHR (Legal)	Watchtower (5 Docs)	Folktales (75 Docs)
Gemini 3.8 Flash	Commercial LLM	Cloud	54.64	17.25	52.50	56.44	69.28	40.33
Google Translate	Commercial API	Cloud	51.26	15.67	48.06	52.30	67.15	37.54
Khaya AI	Commercial API	Cloud	49.91	12.79	47.74	49.92	65.78	36.22
GLOSS-MT (Ours)	LLM + Offline Tool	4B	49.75	11.45	51.23	50.80	60.75	36.23
NLLB-200 3.3B	Dedicated Open MT	3.3B	47.74	14.15	45.25	52.05	56.89	36.76
NLLB-200 1.3B	Dedicated Open MT	1.3B	47.42	13.68	45.72	50.47	57.46	36.05
NLLB-200 600M	Distilled Open MT	600M	46.71	12.74	44.10	50.17	57.06	35.50
Gemma 4 31B Base	Foundation LLM	31B	45.55	11.93	45.97	41.97	57.57	36.71
MADLAD-400 3B	Dedicated Open MT	3B	37.14	6.71	37.43	36.60	50.66	23.87
Gemma 4 E4B Base	Foundation LLM	4B	12.21	0.94	8.18	3.84	21.86	14.96

---

Key Empirical Findings

1. Outperforms All Open Translation Systems: At just 4 billion parameters, GLOSS-MT beats all open-source models on the benchmark, including Meta's NLLB-200 3.3B (49.75 vs 47.74), NLLB-200 1.3B (49.75 vs 47.42), MADLAD-400 3B (49.75 vs 37.14), and Google's Gemma 4 31B Base (49.75 vs 45.55).
2. Domain Champion on Technical Finance: On national budgetary and fiscal documents (Ministry of Finance, 2021–2023), GLOSS-MT achieves 51.23 chrF2, outperforming Google Translate (48.06), Khaya AI (47.74), NLLB-200 3.3B (45.25), and Gemma 4 31B (45.97), placing second only to Gemini 3.8 Flash (52.50).
3. Statutory & Legal Rights (UDHR): Achieves 50.80 chrF2, outperforming Khaya AI (49.92) and completely eliminating the severe hallucination and verbose repetition exhibited by base LLMs (41.97 on Gemma 31B Base; 3.84 on Gemma 4B Base).
4. Religious Literature (Watchtower): Reaches 60.75 chrF2 and 18.92 BLEU, outperforming Gemma 4 31B Base (57.57) and all NLLB models by +3 to +4 chrF points.
5. Authentic Oral Folklore (Akan Folktales): Scores 36.23 chrF2 across 75 historical narrative stories (Rattray 1930) with exact sentence-bounded length fidelity (length ratio 1.19), matching Khaya AI (36.22) and Gemma 4 31B Base (36.71).

---

How It Works: Tool-Calling + Reasoning Trajectory

GLOSS-MT executes translation through an explicit multi-step reasoning trajectory:

Input English Text
       │
       ▼
[1] Tool Call: <|tool_call>call:lookup_twi_translation{phrases: ["court", "decision"]}
       │
       ▼
[2] Offline Dictionary Lookup: 545,633 bilingual entries (no network required)
       │
       ▼
[3] Linguistic Reasoning: <channel|>thought Analysis of syntax, concord, and idioms...
       │
       ▼
[4] Final Translation: Accurate Asante Twi synthesis

---

Quick Start & Usage

1. In-Process Python (with transformers)

Install the official package:

pip install git+https://github.com/GhanaOpenAI/gloss-mt.git

Run in Python:

from twi_translate import TwiTranslator, TransformersBackend

# Load model locally
translator = TwiTranslator(TransformersBackend("ghanaopenai/gloss-mt"))

result = translator.translate("The court announced the decision yesterday.")
print("Translation:", result["translation"])
print("Reasoning:  ", result["reasoning"])
print("Lookups:    ", result["lookups"])

2. High-Throughput Serving (with vLLM)

# Start vLLM inference server
vllm serve ghanaopenai/gloss-mt \
    --served-model-name twi-translator \
    --enable-auto-tool-choice \
    --tool-call-parser gemma4 \
    --reasoning-parser gemma4 \
    --max-model-len 8192 \
    --port 8300

Translate via CLI or client:

twi-translate --base-url http://localhost:8300/v1 --model twi-translator -v "The queen-mother advised the youth to avoid illegal mining."

---

Citation & License

• License: CC-BY-NC-4.0 (Non-commercial research and community deployment)
• Citation: Ghana Open AI & Ghana NLP (2026). GLOSS-MT: Grounding Generative LLMs with External Dictionaries for Low-Resource Machine Translation.