AnViMt

AnViMt — An=English Vi=Vietnamese Mt=Machine Translation — dịch EN → VI cho văn bản thường + kỹ thuật/hệ thống/lập trình/hội thoại.

Export sang ONNX để chạy chỉ với onnxruntime + sentencepiece + numpy (không cần torch/transformers).

Architecture

Base: transformers.BartForConditionalGeneration train từ đầu, không phải pretrained BART.

vocab_size=32000 (SentencePiece BPE, byte_fallback=True)
d_model=512 | encoder_layers=6 | decoder_layers=6
encoder_ffn_dim=2048 | decoder_ffn_dim=2048
encoder_attention_heads=8 | decoder_attention_heads=8
max_position_embeddings=256 | dropout 0.1
tie_word_embeddings=True | is_encoder_decoder=True
pad_token_id=0 (<pad>) | unk_token_id=1 (<unk>) | bos_token_id=2 (<s>) | eos_token_id=3 (</s>) | decoder_start_token_id=2
Params: ~60.79M

EN text
  ↓ spm.model [2] + BPE ids + [3] (pad 0, bos 2, eos 3, max 256)
  ↓
encoder_model.onnx (input_ids, attention_mask → encoder_hidden_states [batch, seq, 512])
  ↓
decoder_model.onnx / decoder_with_past_model.onnx
  (decoder_input_ids [2] + encoder_hidden_states + encoder_attention_mask + past_key_values → logits)
  ↓ autoregressive greedy (argmax) đến eos 3
  ↓ sp.decode → VI text

Files

Export bằng optimum task=seq2seq-lm opset=14:

AnViMt/
├── encoder_model.onnx          # 136M
├── decoder_model.onnx          # 223M
├── decoder_with_past_model.onnx # KV cache (tùy chọn, tăng tốc)
├── spm.model                   # 752K, vocab 32000
└── spm.vocab

Runtime tối thiểu chỉ cần encoder_model.onnx + decoder_model.onnx + spm.model. decoder_with_past_model.onnx dùng khi muốn KV-cache.

Không cần: config.json, pytorch_model.bin, safetensors, torch, transformers.

Installation

pip install onnxruntime sentencepiece numpy
# Termux: apt install python3-onnxruntime python3-sentencepiece python3-numpy

Quick Start

project/
├── main.py
├── encoder_model.onnx
├── decoder_model.onnx
└── spm.model   # (decoder_with_past_model.onnx nếu có)

main.py thực tế (đã test, xử lý đúng BOS/EOS/pad):

import numpy as np, onnxruntime as ort, sentencepiece as spm
sp = spm.SentencePieceProcessor(); sp.load("spm.model")
enc = ort.InferenceSession("encoder_model.onnx", providers=["CPUExecutionProvider"])
dec = ort.InferenceSession("decoder_model.onnx", providers=["CPUExecutionProvider"])
BOS, EOS, PAD = 2, 3, 0
def translate(text, max_len=256):
    ids = [BOS] + sp.encode(text, out_type=int)[:254] + [EOS]
    input_ids = np.array([ids], dtype=np.int64)
    attn = np.ones_like(input_ids)
    enc_hs = enc.run(None, {"input_ids": input_ids, "attention_mask": attn})[0]
    gen = [BOS]
    for _ in range(max_len):
        dec_ids = np.array([gen], dtype=np.int64)
        logits = dec.run(None, {"input_ids": dec_ids, "encoder_hidden_states": enc_hs, "encoder_attention_mask": attn})[0]
        nxt = int(np.argmax(logits[0, -1]))
        gen.append(nxt)
        if nxt == EOS: break
    out = [x for x in gen[1:] if x not in (BOS, EOS, PAD)]
    if EOS in out: out = out[:out.index(EOS)]
    return sp.decode(out)

Training Data

Pretrain 1.13M EN-VI (PhoMT + TED + TECH) → finetune 200k TECH → finetune 27k v3 (hội thoại + kỹ thuật). Tokenizer BPE 32k train trên 400k pairs reservoir sampled, nmt_nfkc, split_digits, byte_fallback.

Limitations

EN → VI only. Câu >254 BPE tokens bị cắt. Thuật ngữ mới có thể giữ nguyên EN.

License

CC BY-NC 4.0 — non-commercial. Author: Phitran21

Downloads last month: -; Downloads are not tracked for this model. How to track

Phitran21
/

AnViMt