v3.0: V2-grammar (g120k) — better grammar/idiom, less repetition (drop-in over v2.2)

README.md

@@ -30,14 +30,30 @@ for web-novel / xianxia content.
 
 - Chinese -> Vietnamese web novel / fiction translation (xianxia, modern, cross-domain).
 - Fast local or server inference where a small model is preferred.
-- This is an experimental release; output should still be reviewed for high-stakes or publication use.
+- Tuned for xianxia / wuxia / classical-flavoured text; on purely modern non-fiction
+  it may occasionally lean to a slightly classical register.
+- Experimental release; review output for high-stakes or publication use.
 
 ## Model Details
 
 - Architecture: Marian seq2seq
 - Parameters: ~37M
 - Tokenizer: SentencePiece source/target tokenizer
-- Suggested decoding: `num_beams=1`, `max_length=512`
+- Suggested decoding: `num_beams=4`, `max_length=512`
+
+## Versions
+
+| Tag | Notes |
+|---|---|
+| **`v3.0`** (current `main`) | Improved classical-grammar + idiom handling; markedly less decoder repetition on short input (no DPO trade-off). |
+| `v2.2` | Previous release. Pin with `revision="v2.2"`. |
+
+Pin a specific version:
+
+```python
+from transformers import AutoModelForSeq2SeqLM
+model = AutoModelForSeq2SeqLM.from_pretrained("DanVP/MoxhiMT-30", revision="v2.2")
+```
 
 ## Quick Start
 
@@ -50,14 +66,13 @@ model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
 
 text = "他抬头看向远处的山门。"
 inputs = tok(text, return_tensors="pt", truncation=True, max_length=512)
-out = model.generate(**inputs, max_length=512, num_beams=1)
+out = model.generate(**inputs, max_length=512, num_beams=4)
 print(tok.decode(out[0], skip_special_tokens=True))
 ```
 
 ## Fast CPU Runtime
 
-A CTranslate2 INT8 export is available in `ct2-int8_float32/` for ~3-5x faster
-CPU inference vs the HuggingFace transformers runtime.
+A CTranslate2 INT8 export is in `ct2-int8_float32/` for ~3-5x faster CPU inference.
 
 ```python
 import ctranslate2
@@ -80,13 +95,14 @@ translator = ctranslate2.Translator(
 ## Training Data
 
 Trained from scratch on a curated Chinese-Vietnamese parallel corpus covering
-xianxia, modern fiction, historical, and cross-domain web-novel content.
+xianxia, modern fiction, historical, and cross-domain web-novel content, with a
+research-grounded layer for idioms and classical-Chinese grammar constructions.
 
 ## Notes
 
-- This model prioritizes speed and small footprint.
+- Prioritizes speed and small footprint.
 - Known hard cases include rare proper nouns and highly domain-specific OOD terminology.
-- For production-style usage, pair with reviewed glossary/guard layers where appropriate.
+- For production usage, pair with reviewed glossary/guard layers where appropriate.
 
 ## License
 

config.json

@@ -1,36 +1,36 @@
-{
-  "activation_dropout": 0.0,
-  "activation_function": "swish",
-  "architectures": [
-    "MarianMTModel"
-  ],
-  "attention_dropout": 0.1,
-  "bos_token_id": 1,
-  "d_model": 448,
-  "decoder_attention_heads": 8,
-  "decoder_ffn_dim": 1792,
-  "decoder_layerdrop": 0.0,
-  "decoder_layers": 2,
-  "decoder_start_token_id": 0,
-  "decoder_vocab_size": 24000,
-  "dropout": 0.1,
-  "dtype": "float32",
-  "encoder_attention_heads": 8,
-  "encoder_ffn_dim": 1792,
-  "encoder_layerdrop": 0.0,
-  "encoder_layers": 8,
-  "eos_token_id": 2,
-  "forced_eos_token_id": 2,
-  "init_std": 0.02,
-  "is_decoder": false,
-  "is_encoder_decoder": true,
-  "max_position_embeddings": 512,
-  "model_type": "marian",
-  "pad_token_id": 0,
-  "scale_embedding": true,
-  "share_encoder_decoder_embeddings": true,
-  "tie_word_embeddings": true,
-  "transformers_version": "5.9.0",
-  "use_cache": true,
-  "vocab_size": 24000
-}
+{
+  "activation_dropout": 0.0,
+  "activation_function": "swish",
+  "architectures": [
+    "MarianMTModel"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "d_model": 448,
+  "decoder_attention_heads": 8,
+  "decoder_ffn_dim": 1792,
+  "decoder_layerdrop": 0.0,
+  "decoder_layers": 2,
+  "decoder_start_token_id": 0,
+  "decoder_vocab_size": 24000,
+  "dropout": 0.1,
+  "dtype": "float32",
+  "encoder_attention_heads": 8,
+  "encoder_ffn_dim": 1792,
+  "encoder_layerdrop": 0.0,
+  "encoder_layers": 8,
+  "eos_token_id": 2,
+  "forced_eos_token_id": 2,
+  "init_std": 0.02,
+  "is_decoder": false,
+  "is_encoder_decoder": true,
+  "max_position_embeddings": 512,
+  "model_type": "marian",
+  "pad_token_id": 0,
+  "scale_embedding": true,
+  "share_encoder_decoder_embeddings": true,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.9.0",
+  "use_cache": true,
+  "vocab_size": 24000
+}

ct2-int8_float32/config.json

@@ -1,10 +1,10 @@
-{
-  "add_source_bos": false,
-  "add_source_eos": false,
-  "bos_token": "<s>",
-  "decoder_start_token": "</s>",
-  "eos_token": "</s>",
-  "layer_norm_epsilon": null,
-  "multi_query_attention": false,
-  "unk_token": "<unk>"
-}
+{
+  "add_source_bos": false,
+  "add_source_eos": false,
+  "bos_token": "<s>",
+  "decoder_start_token": "</s>",
+  "eos_token": "</s>",
+  "layer_norm_epsilon": null,
+  "multi_query_attention": false,
+  "unk_token": "<unk>"
+}

ct2-int8_float32/model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9dff45323096dec7e62a97888b6dd97ea9f8b979014e9c2f623543e2bc716518
+oid sha256:fc5ccad1bf4ee8cd47859130daf2627ae3935bb3c76e017d0c46a563475bc1ef
 size 37896719

ct2-int8_float32/special_tokens_map.json

@@ -1,6 +1,6 @@
 {
-  "bos_token": "<s>",
   "eos_token": "</s>",
+  "unk_token": "<unk>",
   "pad_token": "<pad>",
-  "unk_token": "<unk>"
+  "bos_token": "<s>"
 }

ct2-int8_float32/tokenizer_config.json

@@ -1,50 +1,50 @@
-{
-  "added_tokens_decoder": {
-    "0": {
-      "content": "<pad>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "1": {
-      "content": "<s>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "2": {
-      "content": "</s>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "3": {
-      "content": "<unk>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    }
-  },
-  "backend": "custom",
-  "bos_token": "<s>",
-  "clean_up_tokenization_spaces": false,
-  "eos_token": "</s>",
-  "is_local": true,
-  "local_files_only": false,
-  "model_max_length": 512,
-  "pad_token": "<pad>",
-  "separate_vocabs": false,
-  "source_lang": null,
-  "sp_model_kwargs": {},
-  "target_lang": null,
-  "tokenizer_class": "MarianTokenizer",
-  "unk_token": "<unk>"
-}
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "backend": "custom",
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "separate_vocabs": false,
+  "source_lang": null,
+  "sp_model_kwargs": {},
+  "target_lang": null,
+  "tokenizer_class": "MarianTokenizer",
+  "unk_token": "<unk>"
+}

generation_config.json

@@ -1,12 +1,12 @@
-{
-  "_from_model_config": true,
-  "bos_token_id": 1,
-  "decoder_start_token_id": 0,
-  "eos_token_id": 2,
-  "forced_eos_token_id": 2,
-  "output_attentions": false,
-  "output_hidden_states": false,
-  "pad_token_id": 0,
-  "transformers_version": "5.9.0",
-  "use_cache": true
-}
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "decoder_start_token_id": 0,
+  "eos_token_id": 2,
+  "forced_eos_token_id": 2,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "pad_token_id": 0,
+  "transformers_version": "5.9.0",
+  "use_cache": true
+}

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:342d4a26dba83c8ee54d79cf93d5363ee7a399fdc83906711d95f941137952ad
+oid sha256:a489a6f2a5ab6a6493c05e711bc1bea04804b6c27c5c019a710caca3e35386da
 size 146139424

special_tokens_map.json

@@ -1,6 +1,6 @@
 {
-  "bos_token": "<s>",
   "eos_token": "</s>",
+  "unk_token": "<unk>",
   "pad_token": "<pad>",
-  "unk_token": "<unk>"
+  "bos_token": "<s>"
 }

tokenizer_config.json

@@ -1,50 +1,50 @@
-{
-  "added_tokens_decoder": {
-    "0": {
-      "content": "<pad>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "1": {
-      "content": "<s>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "2": {
-      "content": "</s>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "3": {
-      "content": "<unk>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    }
-  },
-  "backend": "custom",
-  "bos_token": "<s>",
-  "clean_up_tokenization_spaces": false,
-  "eos_token": "</s>",
-  "is_local": true,
-  "local_files_only": false,
-  "model_max_length": 512,
-  "pad_token": "<pad>",
-  "separate_vocabs": false,
-  "source_lang": null,
-  "sp_model_kwargs": {},
-  "target_lang": null,
-  "tokenizer_class": "MarianTokenizer",
-  "unk_token": "<unk>"
-}
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "backend": "custom",
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "is_local": true,
+  "local_files_only": false,
+  "model_max_length": 512,
+  "pad_token": "<pad>",
+  "separate_vocabs": false,
+  "source_lang": null,
+  "sp_model_kwargs": {},
+  "target_lang": null,
+  "tokenizer_class": "MarianTokenizer",
+  "unk_token": "<unk>"
+}