AfriNLP/AfriNLLB-12enc-12dec-full-ft-kd

AfriNLLB

AfriNLLB is a series of efficient multilingual open-source models for African languages.

Model Description

AfriNLLB-12enc-12dec-full-ft-kd is a fully fine-tuned version of NLLB-200 600M on both the AfriNLLB-train and AfriNLLB-train-distilled datasets, without any pruning.

Knowledge distillation: To improve the quality of our models, we employ sequence-level knowledge distillation, where the student model is fine-tuned on a combination of authentic data and synthetic data generated by the teacher model for the same training dataset. In this case, the teacher model is the NLLB-200 3.3B baseline, while the students are the NLLB-200 600M baseline and the the pruned models derived from our fine-tuned version. After generating the data, we filter it by removing duplicates (exact matches in the target side of the authentic data), and we follow a filtering pipeline similar to what we use for processing the original training data.

Supported Languages

AfriNLLB supports 15 language pairs (30 translation directions), including Swahili, Hausa, Yoruba, Amharic, Somali, Zulu, Lingala, Afrikaans, Wolof, and Egyptian Arabic, as well as other African Union official languages such as Arabic (MSA), French, Portuguese, and Spanish. Our training data covers bidirectional translation between English and 13 languages, and between French and two languages (Lingala and Wolof).

Model Details

• Model type: Translation
• Base model: facebook/nllb-200-distilled-600M
• Training data: AfriNLLB-train + AfriNLLB-train-distilled
• Test data: facebook/flores
• Architecture: 12 encoder layers, 12 decoder layers
• Pruning: No

Evaluation Results

We benchmark AfriNLLB models, with diverse iterative layer pruning configurations, against the NLLB baseline across 13 language pairs to/from English and 2 pairs to/from French. Results illustrate the trade-off between translation quality (BLEU, chrF++, COMET) and inference efficiency (output throughput tokens/second and time) at various encoder-decoder layer depths, with and without float16 (FP16) quantization. The current model is marked with 🟣 for the non-quantized version and 🟢 for the float16-quantized version.

Lang	Model	Enc	Dec	Params	V.	Quant	BLEU ↑	chrF++ ↑	COMET ↑	Throughput ↑	Time ↓
xx-en	NLLB	12	12	600M		n/a	33.81	56.22	71.11	1469.96	21.02
						FP16	33.80	56.22	71.13	2834.69	10.92
	AfriNLLB	12	12	600M	🟣	n/a	35.15	57.61	71.87	1530.94	20.39
					🟢	FP16	35.10	57.61	71.87	2808.90	11.15
	AfriNLLB	12	8	548M		n/a	34.01	56.98	71.20	1807.61	17.38
						FP16	34.05	56.99	71.19	3513.32	8.96
	AfriNLLB	12	6	514M		n/a	33.35	56.48	70.79	2028.18	15.41
						FP16	33.32	56.45	70.79	4000.25	7.82
	AfriNLLB	12	4	481M		n/a	32.03	55.62	69.71	2257.03	13.77
						FP16	32.01	55.60	69.71	4589.42	6.79
	AfriNLLB	8	8	498M		n/a	30.89	54.32	68.08	1852.13	17.05
						FP16	30.86	54.30	68.08	3550.50	8.91
en-xx	NLLB	12	12	600M		n/a	22.70	47.89	69.36	1530.10	28.09
						FP16	22.68	47.88	69.38	2898.38	15.33
	AfriNLLB	12	12	600M	🟣	n/a	24.28	49.97	70.91	1610.23	26.98
					🟢	FP16	24.14	49.84	70.90	2811.34	18.82
	AfriNLLB	12	8	548M		n/a	24.17	50.05	70.37	1946.61	22.51
						FP16	24.15	50.06	70.41	3732.72	11.98
	AfriNLLB	12	6	514M		n/a	23.48	49.34	68.98	2265.87	18.50
						FP16	23.49	49.35	69.00	4428.68	9.65
	AfriNLLB	12	4	481M		n/a	21.77	47.80	65.68	2489.35	17.31
						FP16	21.77	47.81	65.68	4954.62	9.09
	AfriNLLB	8	8	498M		n/a	23.59	49.64	69.90	2015.53	21.34
						FP16	23.58	49.63	69.88	3851.13	11.34
xx-fr	NLLB	12	12	600M		n/a	16.41	38.83	17.34	1475.48	26.46
						FP16	16.33	38.83	17.23	2850.66	13.71
	AfriNLLB	12	12	600M	🟣	n/a	17.91	40.45	18.47	1524.32	26.12
					🟢	FP16	17.83	40.42	18.37	2749.45	14.68
	AfriNLLB	12	8	548M		n/a	17.43	40.21	14.52	1845.09	21.61
						FP16	17.38	40.18	14.53	3569.23	11.17
	AfriNLLB	12	6	514M		n/a	16.52	39.44	11.78	2044.27	19.21
						FP16	16.54	39.42	11.68	3953.51	9.92
	AfriNLLB	12	4	481M		n/a	14.96	38.21	5.67	2340.99	16.77
						FP16	14.90	38.17	5.71	4766.12	8.24
	AfriNLLB	8	8	498M		n/a	14.42	37.05	3.14	1866.26	21.84
						FP16	14.34	36.97	3.14	3448.51	11.83
fr-xx	NLLB	12	12	600M		n/a	9.44	33.42	19.25	1047.18	49.92
						FP16	9.52	33.40	19.38	1920.41	29.05
	AfriNLLB	12	12	600M	🟣	n/a	10.98	35.68	21.33	1081.84	51.56
					🟢	FP16	10.48	35.05	21.49	1700.25	51.31
	AfriNLLB	12	8	548M		n/a	10.20	35.21	20.04	1261.66	49.91
						FP16	10.11	35.13	20.03	2313.85	31.15
	AfriNLLB	12	6	514M		n/a	10.07	35.14	19.83	1416.33	30.89
						FP16	9.99	35.08	19.78	2465.60	18.68
	AfriNLLB	12	4	481M		n/a	7.57	32.42	14.16	1207.06	38.75
						FP16	7.57	32.38	14.29	2069.52	23.25
	AfriNLLB	8	8	498M		n/a	9.75	35.23	20.05	1222.83	45.33
						FP16	9.84	35.31	20.11	2186.73	25.97

How to Use

This model can be used for inference with either CTranslate2 or Transformers. The CTranslate2 models are faster and recommended for production. It has two versions, one with float16 quantization ("ct2-fp16") and one without quantization ("ct2"). You can also try the Transformers version, but it is better suited for fine-tuning For detailed code, please refer to the AfriNLLB repositories on GitHub.

pip3 install ctranslate2 sentencepiece transformers huggingface_hub

CTranslate2 version

import os
import ctranslate2
import sentencepiece as spm
from huggingface_hub import snapshot_download, hf_hub_download

src_lang = "eng_Latn"
tgt_lang = "amh_Ethi"

source_sentences = [
    "How are you doing today?",
    "Africa has a diverse history and beautiful nature.",
]

# Download the CTranslate2 model
model_name = "AfriNLP/AfriNLLB-12enc-12dec-full-ft-kd"

# Use "ct2" for the non-quantized version 
# or "ct2-fp16" for the float16 quantized version
ct2_dir = "ct2-fp16"

model_dir = snapshot_download(
    repo_id=model_name,
    allow_patterns=[f"{ct2_dir}/*"]
)
ct2_model_path = os.path.join(model_dir, ct2_dir)

# Download the SentencePiece BPE model
spm_name = "sentencepiece.bpe.model"
spm_path = os.path.join(ct2_model_path, "sentencepiece.bpe.model")
if not os.path.exists(spm_path):
    print("SP model cannot be found locally. Downloading from the baseline...")
    hf_hub_download(
        repo_id="facebook/nllb-200-distilled-600M", 
        filename=spm_name, 
        local_dir=ct2_model_path
        )

sp = spm.SentencePieceProcessor()
sp.load(spm_path)
translator = ctranslate2.Translator(ct2_model_path, device="cuda")

print(f"Translating to {tgt_lang}..\n")

# Tokenize the source texts
encoded_source = sp.encode_as_pieces(source_sentences)
encoded_source = [[src_lang] + s + ["</s>"] for s in encoded_source]

# Translate
results = translator.translate_batch(
    encoded_source,
    target_prefix = [[tgt_lang]] * len(encoded_source),
    beam_size=5,
    max_decoding_length=256,
)

# Decode the outputs and remove the language tag
translations = []
for res in results:
    tokens = res.hypotheses[0]
    if tokens and tokens[0] == tgt_lang:
       tokens = tokens[1:]
    text = sp.decode_pieces(tokens)
    translations.append(text)

for orig, trans in zip(source_sentences, translations):
    print(f"Source ({src_lang}): {orig}\nTarget ({tgt_lang}): {trans}\n")

Transformers version

import torch
from transformers import AutoModelForSeq2SeqLM, NllbTokenizerFast

src_lang = "eng_Latn"
tgt_lang = "amh_Ethi"

source_sentences = [
    "How are you doing today?",
    "Africa has a diverse history and beautiful nature.",
]

# Load an AfriNLLB model
model_name = "AfriNLP/AfriNLLB-12enc-12dec-full-ft-kd"
model = AutoModelForSeq2SeqLM.from_pretrained(
    model_name, 
    device_map="auto",
    )

# Load the NLLB tokenizer
base_model_name = "facebook/nllb-200-distilled-600M"
tokenizer = NllbTokenizerFast.from_pretrained(
    base_model_name,
    src_lang=src_lang,
    )

print(f"\nUsing device: {model.device}")
print(f"Translating to {tgt_lang}..\n")

# Tokenize the source sentences
inputs = tokenizer(
    source_sentences, 
    return_tensors="pt", 
    padding=True).to(model.device)

forced_bos_token_id = tokenizer.convert_tokens_to_ids(tgt_lang)

# Translate
with torch.inference_mode():
    translated_tokens = model.generate(
        **inputs,
        forced_bos_token_id=forced_bos_token_id,
        max_length=256,
        num_beams=5,
        use_cache=True,
    )

# Decode the outputs and remove the language tag
translations = tokenizer.batch_decode(
    translated_tokens, 
    skip_special_tokens=True
    )

for orig, trans in zip(source_sentences, translations):
    print(f"Source ({src_lang}): {orig}\nTarget ({tgt_lang}): {trans}\n")

Citation

If you use any of AfriNLLB models, datasets, or approaches, please cite the following paper:

@inproceedings{moslem-etal-2026-afrinllb,
    title = "{A}fri{NLLB}: Efficient Translation Models for African Languages",
    author = "Moslem, Yasmin  and
      Wassie, Aman Kassahun  and
      Gizachew, Amanuel",
    booktitle = "Proceedings of the Seventh Workshop on African Natural Language Processing (AfricaNLP)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://openreview.net/forum?id=hVJZNUZBur",
}