Upload folder using huggingface_hub

.ipynb_checkpoints/README-checkpoint.md

@@ -7,27 +7,29 @@ tags:
 license: cc-by-4.0
 datasets:
 - quickmt/quickmt-train.ja-en
+- quickmt/madlad400-en-backtranslated-ja
+- quickmt/newscrawl2024-en-backtranslated-ja
 model-index:
 - name: quickmt-ja-en
   results:
   - task:
-      name: Translation jpn-eng
+      name: Translation jap-eng
       type: translation
       args: jpn-eng
     dataset:
       name: flores101-devtest
       type: flores_101
-      args: jpn_Japn eng_Latn devtest
+      args: jpn_Jpan eng_Latn devtest
     metrics:
-    - name: CHRF
-      type: chrf
-      value: 57.06
     - name: BLEU
       type: bleu
-      value: 27.91 
+      value: 28.87 
+    - name: CHRF
+      type: chrf
+      value: 58.56
     - name: COMET
       type: comet
-      value: 87.29
+      value: 87.83
 ---
 
 
@@ -35,14 +37,26 @@ model-index:
 
 `quickmt-ja-en` is a reasonably fast and reasonably accurate neural machine translation model for translation from `ja` into `en`.
 
+`quickmt` models are roughly 3 times faster for GPU inference than OpusMT models and roughly [40 times](https://huggingface.co/spaces/quickmt/quickmt-vs-libretranslate) faster than [LibreTranslate](https://huggingface.co/spaces/quickmt/quickmt-vs-libretranslate)/[ArgosTranslate](github.com/argosopentech/argos-translate). 
+
+
+## *UPDATED!* 
+
+This model was trained with back-translated data and a slightly different configuration and has improved translation quality!
+
+
+## Try it on our Huggingface Space
+
+Give it a try before downloading here: https://huggingface.co/spaces/quickmt/QuickMT-Demo
+
 
 ## Model Information
 
-* Trained using [`eole`](https://github.com/eole-nlp/eole)
-* 185M parameter transformer 'big' with 8 encoder layers and 2 decoder layers
-* 20k sentencepiece vocabularies
+* Trained using [`eole`](https://github.com/eole-nlp/eole) 
+* 200M parameter seq2seq transformer
+* 32k separate Sentencepiece vocabs
 * Exported for fast inference to [CTranslate2](https://github.com/OpenNMT/CTranslate2) format
-* Training data: https://huggingface.co/datasets/quickmt/quickmt-train.ja-en/tree/main
+* The pytorch model (for use with [`eole`](https://github.com/eole-nlp/eole)) is available in this repository in the `eole-model` folder
 
 See the `eole` model configuration in this repository for further details and the `eole-model` for the raw `eole` (pytorch) model.
 
@@ -55,9 +69,9 @@ Next, install the `quickmt` python library and download the model:
 
 ```bash
 git clone https://github.com/quickmt/quickmt.git
-pip install ./quickmt/
+pip install -e ./quickmt/
 
-quickmt-model-download quickmt/quickmt-ja-en ./quickmt-ja-en
+quickmt-model-download quickmt/quickmt-zh-en ./quickmt-ja-en
 ```
 
 Finally use the model in python:
@@ -68,33 +82,35 @@ from quickmt import Translator
 # Auto-detects GPU, set to "cpu" to force CPU inference
 t = Translator("./quickmt-ja-en/", device="auto")
 
-# Translate - set beam size to 5 for higher quality (but slower speed)
+# Translate - set beam size to 1 for faster speed (but lower quality)
 sample_text = 'ノバスコシア州ハリファックスにあるダルハウジー大学医学部教授でカナダ糖尿病協会の臨床・科学部門の責任者を務めるエフード・ウル博士は、この研究はまだ初期段階にあるとして注意を促しました。'
+
 t(sample_text, beam_size=5)
+```
 
-> 'Dr. Ehud Ulu, a professor of medicine at Dalhousie University in Halifax, Nova Scotia and head of the clinical and scientific division of the Canadian Diabetes Association, cautioned that the study is still in its early stages.'
+> 'Dr Ehud Ulu, professor of medicine at Dalhousie University in Halifax, Nova Scotia and head of the clinical and scientific division of the Canadian Diabetes Association, urged caution as the study is still in its infancy.'
 
+```python
 # Get alternative translations by sampling
 # You can pass any cTranslate2 `translate_batch` arguments
 t([sample_text], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)
-
-> 'Dr Ehud Ul, professor of medicine at the University of Dalhousie’s Halifax, Nova Scotia and head of the clinical and scientific division of the Canadian Diabetes Society, noted the study is in its early stages.'
 ```
 
-The model is in `ctranslate2` format, and the tokenizers are `sentencepiece`, so you can use `ctranslate2` directly instead of through `quickmt`. It is also possible  to get this model to work with e.g. [LibreTranslate](https://libretranslate.com/) which also uses `ctranslate2` and `sentencepiece`.
+> 'Dr Ehud Ulu, Professor of Medicine at Dalhousie University in Halifax, Nova Scotia and head of clinical and scientific departments for the Canadian Diabetes Society, urged caution, saying the research was still in its early stages.'
+
+The model is in `ctranslate2` format, and the tokenizers are `sentencepiece`, so you can use `ctranslate2` directly instead of through `quickmt`. It is also possible  to get this model to work with e.g. [LibreTranslate](https://libretranslate.com/) which also uses `ctranslate2` and `sentencepiece`. A model in safetensors format to be used with `eole` is also provided.
 
 
 ## Metrics
 
-`bleu` and `chrf2` are calculated with [sacrebleu](https://github.com/mjpost/sacrebleu) on the [Flores200 `devtest` test set](https://huggingface.co/datasets/facebook/flores) ("jpn_Jpan"->"eng_Latn"). `comet22` with the [`comet`](https://github.com/Unbabel/COMET) library and the [default model](https://huggingface.co/Unbabel/wmt22-comet-da). "Time (s)" is the time in seconds to translate (using `ctranslate2`) the flores-devtest dataset (1012 sentences) on an RTX 4070s GPU with batch size 32 (faster speed is possible using a large batch size).
+`bleu` and `chrf2` are calculated with [sacrebleu](https://github.com/mjpost/sacrebleu) on the [Flores200 `devtest` test set](https://huggingface.co/datasets/facebook/flores), `comet22` with the [`comet`](https://github.com/Unbabel/COMET) library and the [default model](https://huggingface.co/Unbabel/wmt22-comet-da). "Time (s)" is the time in seconds to translate the flores-devtest dataset (1012 sentences) on an RTX 4070s GPU with batch size 32.
+
 
 |                                  |   bleu |   chrf2 |   comet22 |   Time (s) |
 |:---------------------------------|-------:|--------:|----------:|-----------:|
-| quickmt/quickmt-ja-en            |  27.91 |   57.06 |     87.29 |       1.00 |
+| quickmt/quickmt-ja-en            |  28.87 |   58.56 |     87.83 |       1.10 |
 | Helsink-NLP/opus-mt-ja-en        |  19.22 |   49.15 |     82.64 |       3.54 |
 | facebook/nllb-200-distilled-600M |  24.05 |   53.39 |     85.84 |      22.5  |
 | facebook/nllb-200-distilled-1.3B |  28.4  |   56.96 |     87.47 |      37.15 |
 | facebook/m2m100_418M             |  18.82 |   49.55 |     82.59 |      18.27 |
-| facebook/m2m100_1.2B             |  23.32 |   53.46 |     85.43 |      35.44 |
-
-`quickmt-ja-en` is the fastest and nearly as high-quality as facebook/nllb-200-distilled-1.3B.
+| facebook/m2m100_1.2B             |  23.32 |   53.46 |     85.43 |      35.44 |

.ipynb_checkpoints/eole-config-checkpoint.yaml

@@ -0,0 +1,110 @@
+## IO
+save_data: data
+overwrite: True
+seed: 1234
+report_every: 100
+valid_metrics: ["BLEU"]
+tensorboard: true
+tensorboard_log_dir: tensorboard
+
+### Vocab
+src_vocab: ja.eole.vocab
+tgt_vocab: en.eole.vocab
+src_vocab_size: 32000
+tgt_vocab_size: 32000
+vocab_size_multiple: 8
+share_vocab: false
+n_sample: 0
+
+data:
+    corpus_1:
+        path_src: hf://quickmt/quickmt-train.ja-en/ja
+        path_tgt: hf://quickmt/quickmt-train.ja-en/en
+        path_sco: hf://quickmt/quickmt-train.ja-en/sco
+        weight: 2
+    corpus_2:
+        path_src: hf://quickmt/newscrawl2024-en-backtranslated-ja/ja
+        path_tgt: hf://quickmt/newscrawl2024-en-backtranslated-ja/en
+        path_sco: hf://quickmt/newscrawl2024-en-backtranslated-ja/sco
+        weight: 1
+    corpus_3:
+        path_src: hf://quickmt/madlad400-en-backtranslated-ja/ja
+        path_tgt: hf://quickmt/madlad400-en-backtranslated-ja/en
+        path_sco: hf://quickmt/madlad400-en-backtranslated-ja/sco
+        weight: 2
+    valid:
+        path_src: valid.ja
+        path_tgt: valid.en
+
+
+
+transforms: [sentencepiece, filtertoolong]
+transforms_configs:
+  sentencepiece:
+    src_subword_model: "ja.spm.model"
+    tgt_subword_model: "en.spm.model"
+  filtertoolong:
+    src_seq_length: 256
+    tgt_seq_length: 256
+
+training:
+    # Run configuration
+    model_path: quickmt-ja-en-eole-model
+    keep_checkpoint: 4
+    train_steps: 200000
+    save_checkpoint_steps: 5000
+    valid_steps: 5000
+
+    # Train on a single GPU
+    world_size: 1
+    gpu_ranks: [0]
+    batch_type: "tokens"
+    batch_size: 15000
+    valid_batch_size: 2048
+    batch_size_multiple: 8
+    accum_count: [8]
+    accum_steps: [0]
+
+    # Optimizer & Compute
+    compute_dtype: "fp16"
+    optim: "adamw"
+    learning_rate: 3.0
+    warmup_steps: 5000
+    decay_method: "noam"
+    adam_beta2: 0.998
+
+    # Data loading
+    bucket_size: 256000
+    num_workers: 4
+    prefetch_factor: 64
+    
+    # Hyperparams
+    dropout_steps: [0]
+    dropout: [0.1]
+    attention_dropout: [0.1]
+    max_grad_norm: 0
+    label_smoothing: 0.1
+    average_decay: 0.0001
+    param_init_method: xavier_uniform
+    normalization: "tokens"
+
+model:
+    architecture: "transformer"
+    share_embeddings: false
+    share_decoder_embeddings: false
+    add_estimator: false
+    add_ffnbias: true
+    add_qkvbias: false
+    layer_norm: standard
+    mlp_activation_fn: gelu
+    hidden_size: 768
+    encoder:
+        layers: 12
+    decoder:
+        layers: 2
+    heads: 16
+    transformer_ff: 4096
+    embeddings:
+        word_vec_size: 768
+        position_encoding_type: "SinusoidalInterleaved"
+

README.md

@@ -7,27 +7,29 @@ tags:
 license: cc-by-4.0
 datasets:
 - quickmt/quickmt-train.ja-en
+- quickmt/madlad400-en-backtranslated-ja
+- quickmt/newscrawl2024-en-backtranslated-ja
 model-index:
 - name: quickmt-ja-en
   results:
   - task:
-      name: Translation jpn-eng
+      name: Translation jap-eng
       type: translation
       args: jpn-eng
     dataset:
       name: flores101-devtest
       type: flores_101
-      args: jpn_Japn eng_Latn devtest
+      args: jpn_Jpan eng_Latn devtest
     metrics:
-    - name: CHRF
-      type: chrf
-      value: 57.06
     - name: BLEU
       type: bleu
-      value: 27.91 
+      value: 28.87 
+    - name: CHRF
+      type: chrf
+      value: 58.56
     - name: COMET
       type: comet
-      value: 87.29
+      value: 87.83
 ---
 
 
@@ -35,14 +37,26 @@ model-index:
 
 `quickmt-ja-en` is a reasonably fast and reasonably accurate neural machine translation model for translation from `ja` into `en`.
 
+`quickmt` models are roughly 3 times faster for GPU inference than OpusMT models and roughly [40 times](https://huggingface.co/spaces/quickmt/quickmt-vs-libretranslate) faster than [LibreTranslate](https://huggingface.co/spaces/quickmt/quickmt-vs-libretranslate)/[ArgosTranslate](github.com/argosopentech/argos-translate). 
+
+
+## *UPDATED!* 
+
+This model was trained with back-translated data and a slightly different configuration and has improved translation quality!
+
+
+## Try it on our Huggingface Space
+
+Give it a try before downloading here: https://huggingface.co/spaces/quickmt/QuickMT-Demo
+
 
 ## Model Information
 
-* Trained using [`eole`](https://github.com/eole-nlp/eole)
-* 185M parameter transformer 'big' with 8 encoder layers and 2 decoder layers
-* 20k sentencepiece vocabularies
+* Trained using [`eole`](https://github.com/eole-nlp/eole) 
+* 200M parameter seq2seq transformer
+* 32k separate Sentencepiece vocabs
 * Exported for fast inference to [CTranslate2](https://github.com/OpenNMT/CTranslate2) format
-* Training data: https://huggingface.co/datasets/quickmt/quickmt-train.ja-en/tree/main
+* The pytorch model (for use with [`eole`](https://github.com/eole-nlp/eole)) is available in this repository in the `eole-model` folder
 
 See the `eole` model configuration in this repository for further details and the `eole-model` for the raw `eole` (pytorch) model.
 
@@ -55,9 +69,9 @@ Next, install the `quickmt` python library and download the model:
 
 ```bash
 git clone https://github.com/quickmt/quickmt.git
-pip install ./quickmt/
+pip install -e ./quickmt/
 
-quickmt-model-download quickmt/quickmt-ja-en ./quickmt-ja-en
+quickmt-model-download quickmt/quickmt-zh-en ./quickmt-ja-en
 ```
 
 Finally use the model in python:
@@ -68,33 +82,35 @@ from quickmt import Translator
 # Auto-detects GPU, set to "cpu" to force CPU inference
 t = Translator("./quickmt-ja-en/", device="auto")
 
-# Translate - set beam size to 5 for higher quality (but slower speed)
+# Translate - set beam size to 1 for faster speed (but lower quality)
 sample_text = 'ノバスコシア州ハリファックスにあるダルハウジー大学医学部教授でカナダ糖尿病協会の臨床・科学部門の責任者を務めるエフード・ウル博士は、この研究はまだ初期段階にあるとして注意を促しました。'
+
 t(sample_text, beam_size=5)
+```
 
-> 'Dr. Ehud Ulu, a professor of medicine at Dalhousie University in Halifax, Nova Scotia and head of the clinical and scientific division of the Canadian Diabetes Association, cautioned that the study is still in its early stages.'
+> 'Dr Ehud Ulu, professor of medicine at Dalhousie University in Halifax, Nova Scotia and head of the clinical and scientific division of the Canadian Diabetes Association, urged caution as the study is still in its infancy.'
 
+```python
 # Get alternative translations by sampling
 # You can pass any cTranslate2 `translate_batch` arguments
 t([sample_text], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)
-
-> 'Dr Ehud Ul, professor of medicine at the University of Dalhousie’s Halifax, Nova Scotia and head of the clinical and scientific division of the Canadian Diabetes Society, noted the study is in its early stages.'
 ```
 
-The model is in `ctranslate2` format, and the tokenizers are `sentencepiece`, so you can use `ctranslate2` directly instead of through `quickmt`. It is also possible  to get this model to work with e.g. [LibreTranslate](https://libretranslate.com/) which also uses `ctranslate2` and `sentencepiece`.
+> 'Dr Ehud Ulu, Professor of Medicine at Dalhousie University in Halifax, Nova Scotia and head of clinical and scientific departments for the Canadian Diabetes Society, urged caution, saying the research was still in its early stages.'
+
+The model is in `ctranslate2` format, and the tokenizers are `sentencepiece`, so you can use `ctranslate2` directly instead of through `quickmt`. It is also possible  to get this model to work with e.g. [LibreTranslate](https://libretranslate.com/) which also uses `ctranslate2` and `sentencepiece`. A model in safetensors format to be used with `eole` is also provided.
 
 
 ## Metrics
 
-`bleu` and `chrf2` are calculated with [sacrebleu](https://github.com/mjpost/sacrebleu) on the [Flores200 `devtest` test set](https://huggingface.co/datasets/facebook/flores) ("jpn_Jpan"->"eng_Latn"). `comet22` with the [`comet`](https://github.com/Unbabel/COMET) library and the [default model](https://huggingface.co/Unbabel/wmt22-comet-da). "Time (s)" is the time in seconds to translate (using `ctranslate2`) the flores-devtest dataset (1012 sentences) on an RTX 4070s GPU with batch size 32 (faster speed is possible using a large batch size).
+`bleu` and `chrf2` are calculated with [sacrebleu](https://github.com/mjpost/sacrebleu) on the [Flores200 `devtest` test set](https://huggingface.co/datasets/facebook/flores), `comet22` with the [`comet`](https://github.com/Unbabel/COMET) library and the [default model](https://huggingface.co/Unbabel/wmt22-comet-da). "Time (s)" is the time in seconds to translate the flores-devtest dataset (1012 sentences) on an RTX 4070s GPU with batch size 32.
+
 
 |                                  |   bleu |   chrf2 |   comet22 |   Time (s) |
 |:---------------------------------|-------:|--------:|----------:|-----------:|
-| quickmt/quickmt-ja-en            |  27.91 |   57.06 |     87.29 |       1.00 |
+| quickmt/quickmt-ja-en            |  28.87 |   58.56 |     87.83 |       1.10 |
 | Helsink-NLP/opus-mt-ja-en        |  19.22 |   49.15 |     82.64 |       3.54 |
 | facebook/nllb-200-distilled-600M |  24.05 |   53.39 |     85.84 |      22.5  |
 | facebook/nllb-200-distilled-1.3B |  28.4  |   56.96 |     87.47 |      37.15 |
 | facebook/m2m100_418M             |  18.82 |   49.55 |     82.59 |      18.27 |
-| facebook/m2m100_1.2B             |  23.32 |   53.46 |     85.43 |      35.44 |
-
-`quickmt-ja-en` is the fastest and nearly as high-quality as facebook/nllb-200-distilled-1.3B.
+| facebook/m2m100_1.2B             |  23.32 |   53.46 |     85.43 |      35.44 |

eole-config.yaml

@@ -10,8 +10,8 @@ tensorboard_log_dir: tensorboard
 ### Vocab
 src_vocab: ja.eole.vocab
 tgt_vocab: en.eole.vocab
-src_vocab_size: 20000
-tgt_vocab_size: 20000
+src_vocab_size: 32000
+tgt_vocab_size: 32000
 vocab_size_multiple: 8
 share_vocab: false
 n_sample: 0
@@ -21,10 +21,23 @@ data:
         path_src: hf://quickmt/quickmt-train.ja-en/ja
         path_tgt: hf://quickmt/quickmt-train.ja-en/en
         path_sco: hf://quickmt/quickmt-train.ja-en/sco
+        weight: 2
+    corpus_2:
+        path_src: hf://quickmt/newscrawl2024-en-backtranslated-ja/ja
+        path_tgt: hf://quickmt/newscrawl2024-en-backtranslated-ja/en
+        path_sco: hf://quickmt/newscrawl2024-en-backtranslated-ja/sco
+        weight: 1
+    corpus_3:
+        path_src: hf://quickmt/madlad400-en-backtranslated-ja/ja
+        path_tgt: hf://quickmt/madlad400-en-backtranslated-ja/en
+        path_sco: hf://quickmt/madlad400-en-backtranslated-ja/sco
+        weight: 2
     valid:
         path_src: valid.ja
         path_tgt: valid.en
 
+
+
 transforms: [sentencepiece, filtertoolong]
 transforms_configs:
   sentencepiece:
@@ -36,20 +49,18 @@ transforms_configs:
 
 training:
     # Run configuration
-    model_path: eole-model
+    model_path: quickmt-ja-en-eole-model
     keep_checkpoint: 4
-    train_steps: 108000
-    save_checkpoint_steps: 2000
-    valid_steps: 2000
-    
+    train_steps: 200000
+    save_checkpoint_steps: 5000
+    valid_steps: 5000
+
     # Train on a single GPU
     world_size: 1
     gpu_ranks: [0]
-
-    # Batching 10240
     batch_type: "tokens"
-    batch_size: 8000
-    valid_batch_size: 4096
+    batch_size: 15000
+    valid_batch_size: 2048
     batch_size_multiple: 8
     accum_count: [8]
     accum_steps: [0]
@@ -57,15 +68,15 @@ training:
     # Optimizer & Compute
     compute_dtype: "fp16"
     optim: "adamw"
-    learning_rate: 2.0
-    warmup_steps: 4000
+    learning_rate: 3.0
+    warmup_steps: 5000
     decay_method: "noam"
     adam_beta2: 0.998
 
     # Data loading
-    bucket_size: 128000
+    bucket_size: 256000
     num_workers: 4
-    prefetch_factor: 32
+    prefetch_factor: 64
     
     # Hyperparams
     dropout_steps: [0]
@@ -81,14 +92,19 @@ model:
     architecture: "transformer"
     share_embeddings: false
     share_decoder_embeddings: false
-    hidden_size: 1024
+    add_estimator: false
+    add_ffnbias: true
+    add_qkvbias: false
+    layer_norm: standard
+    mlp_activation_fn: gelu
+    hidden_size: 768
     encoder:
-        layers: 8
+        layers: 12
     decoder:
         layers: 2
-    heads: 8
+    heads: 16
     transformer_ff: 4096
     embeddings:
-        word_vec_size: 1024
+        word_vec_size: 768
         position_encoding_type: "SinusoidalInterleaved"
 

eole-model/config.json

@@ -1,108 +1,121 @@
 {
-  "n_sample": 0,
-  "overwrite": true,
+  "src_vocab": "ja.eole.vocab",
   "save_data": "data",
-  "tgt_vocab_size": 20000,
-  "seed": 1234,
-  "share_vocab": false,
+  "valid_metrics": [
+    "BLEU"
+  ],
   "tensorboard_log_dir": "tensorboard",
   "report_every": 100,
-  "tensorboard": true,
-  "tgt_vocab": "en.eole.vocab",
   "vocab_size_multiple": 8,
-  "src_vocab": "ja.eole.vocab",
-  "src_vocab_size": 20000,
-  "tensorboard_log_dir_dated": "tensorboard/Apr-15_05-55-59",
+  "seed": 1234,
   "transforms": [
     "sentencepiece",
     "filtertoolong"
   ],
-  "valid_metrics": [
-    "BLEU"
-  ],
+  "overwrite": true,
+  "src_vocab_size": 32000,
+  "tgt_vocab": "en.eole.vocab",
+  "share_vocab": false,
+  "tensorboard": true,
+  "n_sample": 0,
+  "tgt_vocab_size": 32000,
+  "tensorboard_log_dir_dated": "tensorboard/Dec-24_20-29-50",
   "training": {
-    "save_checkpoint_steps": 2000,
-    "max_grad_norm": 0.0,
-    "dropout": [
+    "batch_type": "tokens",
+    "param_init_method": "xavier_uniform",
+    "batch_size_multiple": 8,
+    "learning_rate": 3.0,
+    "attention_dropout": [
       0.1
     ],
-    "dropout_steps": [
+    "accum_steps": [
       0
     ],
-    "label_smoothing": 0.1,
+    "batch_size": 15000,
+    "model_path": "quickmt-ja-en-eole-model",
     "keep_checkpoint": 4,
-    "prefetch_factor": 32,
-    "valid_steps": 2000,
-    "optim": "adamw",
+    "adam_beta2": 0.998,
     "gpu_ranks": [
       0
     ],
-    "bucket_size": 128000,
-    "batch_size": 8000,
     "average_decay": 0.0001,
-    "compute_dtype": "torch.float16",
-    "param_init_method": "xavier_uniform",
-    "decay_method": "noam",
-    "num_workers": 0,
-    "warmup_steps": 4000,
-    "attention_dropout": [
+    "warmup_steps": 5000,
+    "valid_steps": 5000,
+    "dropout": [
       0.1
     ],
-    "model_path": "quickmt-ja-en-eole-model",
-    "accum_count": [
-      8
+    "dropout_steps": [
+      0
     ],
-    "batch_size_multiple": 8,
+    "prefetch_factor": 64,
+    "max_grad_norm": 0.0,
+    "world_size": 1,
+    "compute_dtype": "torch.float16",
     "train_steps": 200000,
-    "accum_steps": [
-      0
+    "accum_count": [
+      8
     ],
-    "adam_beta2": 0.998,
-    "learning_rate": 2.0,
+    "num_workers": 0,
     "normalization": "tokens",
-    "valid_batch_size": 4096,
-    "world_size": 1,
-    "batch_type": "tokens"
+    "decay_method": "noam",
+    "optim": "adamw",
+    "valid_batch_size": 2048,
+    "bucket_size": 256000,
+    "label_smoothing": 0.1,
+    "save_checkpoint_steps": 5000
   },
   "model": {
+    "layer_norm": "standard",
+    "hidden_size": 768,
+    "add_estimator": false,
+    "mlp_activation_fn": "gelu",
+    "share_embeddings": false,
+    "heads": 16,
     "transformer_ff": 4096,
+    "add_qkvbias": false,
     "share_decoder_embeddings": false,
-    "architecture": "transformer",
     "position_encoding_type": "SinusoidalInterleaved",
-    "share_embeddings": false,
-    "heads": 8,
-    "hidden_size": 1024,
+    "add_ffnbias": true,
+    "architecture": "transformer",
     "encoder": {
-      "layers": 8,
-      "transformer_ff": 4096,
-      "position_encoding_type": "SinusoidalInterleaved",
-      "src_word_vec_size": 1024,
+      "src_word_vec_size": 768,
       "n_positions": null,
+      "layer_norm": "standard",
       "encoder_type": "transformer",
-      "heads": 8,
-      "hidden_size": 1024
+      "hidden_size": 768,
+      "mlp_activation_fn": "gelu",
+      "heads": 16,
+      "transformer_ff": 4096,
+      "position_encoding_type": "SinusoidalInterleaved",
+      "add_ffnbias": true,
+      "add_qkvbias": false,
+      "layers": 12
     },
     "embeddings": {
-      "tgt_word_vec_size": 1024,
-      "position_encoding_type": "SinusoidalInterleaved",
-      "word_vec_size": 1024,
-      "src_word_vec_size": 1024
+      "word_vec_size": 768,
+      "src_word_vec_size": 768,
+      "tgt_word_vec_size": 768,
+      "position_encoding_type": "SinusoidalInterleaved"
     },
     "decoder": {
+      "n_positions": null,
+      "layer_norm": "standard",
+      "hidden_size": 768,
+      "mlp_activation_fn": "gelu",
+      "tgt_word_vec_size": 768,
       "layers": 2,
+      "heads": 16,
       "transformer_ff": 4096,
       "position_encoding_type": "SinusoidalInterleaved",
-      "n_positions": null,
-      "decoder_type": "transformer",
-      "tgt_word_vec_size": 1024,
-      "heads": 8,
-      "hidden_size": 1024
+      "add_ffnbias": true,
+      "add_qkvbias": false,
+      "decoder_type": "transformer"
     }
   },
   "transforms_configs": {
     "sentencepiece": {
-      "src_subword_model": "${MODEL_PATH}/ja.spm.model",
-      "tgt_subword_model": "${MODEL_PATH}/en.spm.model"
+      "tgt_subword_model": "${MODEL_PATH}/en.spm.model",
+      "src_subword_model": "${MODEL_PATH}/ja.spm.model"
     },
     "filtertoolong": {
       "tgt_seq_length": 256,
@@ -111,22 +124,46 @@
   },
   "data": {
     "corpus_1": {
-      "path_tgt": "en.txt",
+      "path_tgt": "hf://quickmt/quickmt-train.ja-en/en",
+      "path_align": null,
+      "path_src": "hf://quickmt/quickmt-train.ja-en/ja",
+      "transforms": [
+        "sentencepiece",
+        "filtertoolong"
+      ],
+      "weight": 2,
+      "path_sco": "hf://quickmt/quickmt-train.ja-en/sco"
+    },
+    "corpus_2": {
+      "path_tgt": "hf://quickmt/newscrawl2024-en-backtranslated-ja/en",
+      "path_align": null,
+      "path_src": "hf://quickmt/newscrawl2024-en-backtranslated-ja/ja",
+      "transforms": [
+        "sentencepiece",
+        "filtertoolong"
+      ],
+      "weight": 1,
+      "path_sco": "hf://quickmt/newscrawl2024-en-backtranslated-ja/sco"
+    },
+    "corpus_3": {
+      "path_tgt": "hf://quickmt/madlad400-en-backtranslated-ja/en",
       "path_align": null,
+      "path_src": "hf://quickmt/madlad400-en-backtranslated-ja/ja",
       "transforms": [
         "sentencepiece",
         "filtertoolong"
       ],
-      "path_src": "ja.txt"
+      "weight": 2,
+      "path_sco": "hf://quickmt/madlad400-en-backtranslated-ja/sco"
     },
     "valid": {
-      "path_tgt": "valid.en",
       "path_align": null,
+      "path_src": "valid.ja",
       "transforms": [
         "sentencepiece",
         "filtertoolong"
       ],
-      "path_src": "valid.ja"
+      "path_tgt": "valid.en"
     }
   }
 }

eole-model/en.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:01dc857df9df5987bec61812d1fbc2e78ad530346c09fe3ffaf27184358ab8fd
-size 583983
+oid sha256:d2e58f48c0d030d6b4e6b183ebca7068704ef9cce640f090abfea4ac9ea11da0
+size 797948

eole-model/ja.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cbd449d7e92d850f6d36efeafb40cea3a8468f55db0ca751ee4ece0ceb70d19b
-size 583133
+oid sha256:438e5f7fd498d76b057ca8b6759470b82eb3855b81a67718face917407b31ff5
+size 816774

eole-model/model.00.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f226e709170b27c12510b5b186f305fcc9ddb3dbf4ae3ea1066f8ff4632f7de1
-size 823882912
+oid sha256:850944f80d706c273021d0dc80a82d05536672532122dccccaeb63065d74ace3
+size 829569112

model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4f9906f44709dab21e2bd168cb19e4379326d4b8e92513dc9b4e5ff4af6cf323
-size 401699775
+oid sha256:d11276f68986d951edc1e5b4b634e00f1f9c493eb14519598be975630965eb47
+size 407101843

src.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cbd449d7e92d850f6d36efeafb40cea3a8468f55db0ca751ee4ece0ceb70d19b
-size 583133
+oid sha256:438e5f7fd498d76b057ca8b6759470b82eb3855b81a67718face917407b31ff5
+size 816774

tgt.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:01dc857df9df5987bec61812d1fbc2e78ad530346c09fe3ffaf27184358ab8fd
-size 583983
+oid sha256:d2e58f48c0d030d6b4e6b183ebca7068704ef9cce640f090abfea4ac9ea11da0
+size 797948