import gradio as gr import torch from transformers import AutoTokenizer, AutoModelForSeq2SeqLM from peft import PeftModel # Multi adaptor config BASE_MODEL_ID = "facebook/nllb-200-distilled-600M" EN_TO_KIK = "KnoxDevelopers/English-to-Kikuyu-language-translation-LoRA" EN_TO_LUO = "KnoxDevelopers/English-to-Luo-language-translation-LoRA" EN_TO_KAM = "KnoxDevelopers/nllb-200-English_to_Kikamba-lang-translation-QLoRA" EN_TO_DAV = "KnoxDevelopers/English-to-Taita-language-translation-LoRA" EN_TO_GUZ = "KnoxDevelopers/English-to-Kisii-language-translation-LoRA" EN_TO_MER = "KnoxDevelopers/English-to-Meru-language-translation-LoRA" EN_TO_KLN = "KnoxDevelopers/English-to-Kalenjin-language-translation-LoRA" EN_TO_EBU = "KnoxDevelopers/English-to-Embu-language-translation-LoRA" EN_TO_MAS = "KnoxDevelopers/English-to-Maasai-language-translation-LoRA" print("[*] Loading base tokenizer and model...") tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID) base_model = AutoModelForSeq2SeqLM.from_pretrained( BASE_MODEL_ID, tie_word_embeddings=False ) base_model.resize_token_embeddings(256205) if hasattr(base_model, "config"): base_model.config.ensure_weight_tying = True print("[*] adding adapters...") model = PeftModel.from_pretrained(base_model, EN_TO_KIK, adapter_name="en_to_kik") model.load_adapter(EN_TO_LUO, adapter_name="en_to_luo") model.load_adapter(EN_TO_KAM, adapter_name="en_to_kam") model.load_adapter(EN_TO_DAV, adapter_name="en_to_dav") model.load_adapter(EN_TO_GUZ, adapter_name="en_to_guz") model.load_adapter(EN_TO_MER, adapter_name="en_to_mer") model.load_adapter(EN_TO_KLN, adapter_name="en_to_kln") model.load_adapter(EN_TO_EBU, adapter_name="en_to_ebu") model.load_adapter(EN_TO_MAS, adapter_name="en_to_mas") device = "cuda" if torch.cuda.is_available() else "cpu" model.to(device) model.eval() def get_safe_token_id(lang_code, fallback_code="kik_Latn"): token_id = tokenizer.convert_tokens_to_ids(lang_code) max_vocab_limit = base_model.config.vocab_size if token_id is None or token_id == tokenizer.unk_token_id or token_id >= max_vocab_limit: token_id = tokenizer.convert_tokens_to_ids(fallback_code) return token_id MAX_CHARACTER_LIMIT = 400 BANNED_WORDS = ["offensive_placeholder"] def programmatic_guardrails(text): text_clean = text.strip() if not text_clean: return False, "⚠️ Please enter a valid sentence to translate." if len(text_clean) > MAX_CHARACTER_LIMIT: return False, f"⚠️ Input exceeds the safety limit of {MAX_CHARACTER_LIMIT} characters." for word in BANNED_WORDS: if word in text_clean.lower(): return False, "🛑 Safety Violation: Input violates our translation safety policy." return True, text_clean def translate_interface(source_text, direction): is_safe, sanitized_input = programmatic_guardrails(source_text) if not is_safe: return sanitized_input try: tokenizer.src_lang = "eng_Latn" inputs = tokenizer(sanitized_input, return_tensors="pt").to(device) if direction == "English to Kikuyu": model.set_adapter("en_to_kik") target_lang_id = get_safe_token_id("kik_Latn") elif direction == "English to Luo": model.set_adapter("en_to_luo") target_lang_id = get_safe_token_id("luo_Latn") elif direction == "English to Kikamba": model.set_adapter("en_to_kam") target_lang_id = get_safe_token_id("kam_Latn") elif direction == "English to Taita": model.set_adapter("en_to_dav") target_lang_id = get_safe_token_id("dav_Latn") elif direction == "English to Kisii": model.set_adapter("en_to_guz") target_lang_id = get_safe_token_id("guz_Latn") elif direction == "English to Meru": model.set_adapter("en_to_mer") target_lang_id = get_safe_token_id("mer_Latn") elif direction == "English to Kalenjin": model.set_adapter("en_to_kln") target_lang_id = get_safe_token_id("kln_Latn") elif direction == "English to Embu": model.set_adapter("en_to_ebu") target_lang_id = get_safe_token_id("ebu_Latn") elif direction == "English to Maasai": model.set_adapter("en_to_mas") target_lang_id = get_safe_token_id("mas_Latn") else: return "Invalid translation direction selected." with torch.no_grad(): generated_tokens = model.generate( **inputs, forced_bos_token_id=target_lang_id, max_length=512, num_beams=4, early_stopping=True ) return tokenizer.decode(generated_tokens[0], skip_special_tokens=True) except Exception as e: return f"An internal error occurred during generation: {str(e)}" custom_theme = gr.themes.Soft( font=[gr.themes.GoogleFont("Arial"), "sans-serif"] ) description_html = """

Multilingual Kenyan Language Translator

Translate English into Kikuyu, Luo, Kikamba, Taita, Kisii, Meru, Kalenjin, Embu and Maasai using fine-tuned Meta NLLB-200 LoRA adapters.

Trained by Knox Systems Developers

""" with gr.Blocks(theme=custom_theme) as demo: gr.HTML(description_html) direction_selector = gr.Radio( choices=[ "English to Kikuyu", "English to Luo", "English to Kikamba", "English to Taita", "English to Kisii", "English to Meru", "English to Kalenjin", "English to Embu", "English to Maasai" ], value="English to Kikuyu", label="Translation Direction" ) with gr.Row(): with gr.Column(): input_box = gr.Textbox( label="Source Input Text", placeholder="Type your English text here...", lines=4 ) submit_btn = gr.Button("Translate Text", variant="primary") with gr.Column(): output_box = gr.Textbox( label="Translated Output", interactive=False, lines=4 ) submit_btn.click( fn=translate_interface, inputs=[input_box, direction_selector], outputs=output_box ) gr.Examples( examples=[ ["The children are playing outside near the tree.", "English to Kikuyu"], ["Where is the market? I need to buy some food.", "English to Luo"], ["The children are playing outside near the tree.", "English to Kikamba"], ["We will meet tomorrow morning.", "English to Taita"], ["Please help me carry this bag.", "English to Kisii"], ["The children are playing outside near the tree.", "English to Meru"], ["Where is the market? I need to buy some food.", "English to Kalenjin"], ["The children are playing outside near the tree.", "English to Embu"], ["Where is the market? I need to buy some food.", "English to Maasai"] ], inputs=[input_box, direction_selector] ) demo.launch()