ghanaopenai/twi-translation-tool-calling

Twi Translation Tool-Calling with Reasoning

Author: Ghana Open AI
Supported by: Ghana NLP

SFT data that teaches an LLM to translate English into Twi (Akan) like a careful translator:

1. look up the key English noun phrases in a dictionary (lookup_twi_translation tool call),
2. read the results (a tool message; phrases that aren't in the dictionary come back empty),
3. reason about grammar, idioms and word choice, using the dictionary terms where they fit,
4. translate.

Each row is a full agent trajectory in the standard chat format with tools:

system -> user -> assistant(tool_calls) -> tool(results) -> assistant(reasoning_content + content)

Columns

For browsing, each row has the trajectory as flat columns, in order:

column	content
english	source text
tool_phrases	phrases the model looks up (the tool-call arguments)
tool_result	what the tool returns (fed back to the model)
reasoning	step-by-step linguistic reasoning, written after the lookup
translation	gold Twi translation

messages + tools hold the same content in the chat format used for training. The reasoning is in the final assistant message's reasoning_content field, separate from content.

Where each part comes from

Part	Source
English text, reasoning, Twi translation	English→Twi rows (pristine, ghana_chat_en2tw) of ghananlpcommunity/twi-english-reasoning-translation. The Twi side is the gold human reference, and the reasoning explains it
Tool-call phrases	the English phrases the reasoning quotes (e.g. 'national progress' is 'ɔman nkɔso'), kept only if they occur in the source text and are noun phrases (spaCy), with articles and possessives removed, at most 4 words
Tool results	those phrases looked up in an offline English→Twi dictionary of 478,822 phrases, built from the Gemini translations in AfriSpeech/shola-machine-translations (gemini-3.6-flash-one, falling back to up to 3 options from gemini-3.6-flash)

Because the tool results come from the same dictionary the deployed tool uses, a model trained on this data receives exactly the kind of tool output it saw in training. Dictionary entries are machine candidates, not verified translations, so the reasoning and gold translations sometimes use a different term. That teaches the model to treat lookups as hints, not ground truth.

Splits

split	examples	with tool call	phrases looked up	found in dictionary
train	15,134	14,431	139,747	41.2%
validation	332	322	3,166	40.6%
test	538	513	4,864	40.8%

Splits are a stable hash of the source row id. About 70% of rows carry Markdown formatting (headings, lists, tables, bold and italic) that the translation preserves.

Example (abridged)

{"messages": [
  {"role": "system", "content": "You are an expert English-to-Twi (Akan) translator. First call the lookup_twi_translation tool ..."},
  {"role": "user", "content": "Translate the following English text into Twi (Akan). Return only the translation with no additional text.\n\n..."},
  {"role": "assistant", "content": "", "tool_calls": [{"id": "call_000123", "type": "function",
    "function": {"name": "lookup_twi_translation", "arguments": {"phrases": ["cocoa heaps", "proper fermentation", "inspection"]}}}]},
  {"role": "tool", "tool_call_id": "call_000123", "name": "lookup_twi_translation",
   "content": "Twi lookup results (2 of 3 phrases found):\n- cocoa heaps -> koko sumie | koko ekuo | koko boaboa ano\n- inspection -> nhwehwɛmu\nNot found (translate from your own knowledge): proper fermentation"},
  {"role": "assistant", "reasoning_content": "1. 'Sɛnea nhwehwɛmu kyerɛ no' translates 'According to research' ...", "content": "Sɛnea nhwehwɛmu kyerɛ no, ..."}
 ],
 "tools": [{"type": "function", "function": {"name": "lookup_twi_translation", "parameters": {"type": "object", "properties": {"phrases": {"type": "array", "items": {"type": "string"}}}, "required": ["phrases"]}}}],
 "id": "...", "source": "pristine", "markdown": "heading,bullet_list", "num_phrases": 3, "num_found": 2}

reasoning_content is the standard field that chat templates (e.g. Gemma 4, Qwen 3) render as the model's thinking. With Gemma 4, render with enable_thinking=True so the thought channel opens after the tool result.

Code

The dataset builder, the offline dictionary, the training script and a ready-to-run deployment runtime (tool-call parsing → dictionary lookup → result fed back → translation, with transformers, vLLM or an HTTP API) are in the accompanying GitHub repository.

Licence

CC-BY-NC-4.0, inherited from both sources. Non-commercial use only.