You can deploy this model and start earning money today!

drrobot9/
nllb-yoruba-farming-finetuned

Name: nllb-yoruba-farming-finetuned
Availability: InStock

safetensors

m2m_100

Views

No views yet

language:

en
yo tags:
translation
nllb
yoruba
agriculture
nigerian-languages
seq2seq license: cc-by-nc-4.0 base_model: facebook/nllb-200-distilled-600M pipeline_tag: translation

NLLB Yoruba Farming Fine-tuned

A fine-tuned version of facebook/nllb-200-distilled-600M specialised for English ↔ Yoruba translation in the Nigerian agricultural domain, built for FarmLingua AI by Kawafarm LTD.

Model Description

This model extends NLLB-200 with domain adaptation on Nigerian farming vocabulary, crop management terminology, livestock care, and agribusiness language. It is designed to produce natural, fluent Yoruba output for farming-related content generated by English-language LLMs.

Training Data

Dataset	Size	Domain
odunola/yoruba-english-pairs	~20k pairs	General Yoruba-English
Custom farm_yoruba farming pairs	233 pairs × 10 oversample = 2,330	Nigerian agriculture
Combined	~22,330 pairs	General + Agricultural

The farming pairs cover: rice, maize, cassava, yam, tomato, plantain, groundnut, cocoa, oil palm, cashew, ginger, catfish, poultry (broilers, layers), pigs, goats, dairy cattle, snail farming, bee farming, rubber tapping, and post-harvest management.

Training Details

Parameter	Value
Base model	facebook/nllb-200-distilled-600M
Training epochs	5
Batch size	32
Learning rate	5e-5
Warmup steps	200
Max sequence length	256
Beam search	4 beams
Hardware	NVIDIA A100
Framework	HuggingFace Transformers

Languages

Code	Language	Direction
`eng_Latn`	English	Source / Target
`yor_Latn`	Yoruba	Source / Target

Usage

python

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

model_id  = "drrobot9/nllb-yoruba-farming-finetuned"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model     = AutoModelForSeq2SeqLM.from_pretrained(model_id).to("cuda")
model.eval()

def translate(text, src_lang="eng_Latn", tgt_lang="yor_Latn"):
    tokenizer.src_lang = src_lang
    inputs = tokenizer(text, return_tensors="pt", truncation=True).to("cuda")
    forced_bos = tokenizer.convert_tokens_to_ids(tgt_lang)
    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            forced_bos_token_id=forced_bos,
            num_beams=4,
            max_new_tokens=256,
        )
    return tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]


# English → Yoruba
print(translate("How do I start a rice farm in Nigeria?"))

# Yoruba → English
print(translate(
    "Bawo ni mo ṣe le bẹrẹ oko iresi ni Naijiria?",
    src_lang="yor_Latn",
    tgt_lang="eng_Latn"
))

Intended Use

This model is a translation component within the FarmLingua AI pipeline:

User input (Yoruba/English)
        ↓
Language detection (facebook/fasttext-language-identification)
        ↓
Translation → English  [this model]
        ↓
Qwen2.5-1.5B-Instruct  (farming reasoning in English)
        ↓
Translation → Yoruba   [this model]
        ↓
User receives answer in their language


## Limitations

- Optimised for **agricultural domain text** — general-purpose translation quality may vary
- Trained on **English ↔ Yoruba** only — does not handle Igbo or Hausa (use base NLLB for those)
- Yoruba tonal diacritics accuracy depends on training data quality
- Not intended for legal, medical, or financial translation

## Built By

**Kawafarm LTD** — *Empowering Nigerian farmers through AI*

> FarmLingua AI was built to help Nigerian farmers access agricultural knowledge in their local languages.
"""

# Write model card to output directory and push
with open(f"{OUTPUT_DIR}/README.md", "w", encoding="utf-8") as f:
    f.write(model_card)

print("Model card written.")

# Push updated README to Hub
api.upload_file(
    path_or_fileobj=f"{OUTPUT_DIR}/README.md",
    path_in_repo="README.md",
    repo_id=REPO_ID,
    repo_type="model",
    commit_message="Add model card",
    token=TOKEN,
)

print(f"Model card pushed → https://huggingface.co/{REPO_ID}")

Run this as a new cell after your upload cell. It writes the README.md locally and pushes it to the Hub in one step.