Upload folder using huggingface_hub

.ipynb_checkpoints/README-checkpoint.md

@@ -1,21 +1,83 @@
+---
+language:
+- en
+- zh
+tags:
+- translation
+license: cc-by-4.0
+datasets:
+- quickmt/quickmt-train.zh-en
+- quickmt/madlad400-en-backtranslated-zh
+- quickmt/newscrawl2024-en-backtranslated-zh
+model-index:
+- name: quickmt-zh-en
+  results:
+  - task:
+      name: Translation zho-eng
+      type: translation
+      args: zho-eng
+    dataset:
+      name: flores101-devtest
+      type: flores_101
+      args: zho_Hans eng_Latn devtest
+    metrics:
+    - name: BLEU
+      type: bleu
+      value: 29.9
+    - name: CHRF
+      type: chrf
+      value: 58.42
+    - name: COMET
+      type: comet
+      value: 86.59
+---
+
+
 # `quickmt-zh-en` Neural Machine Translation Model 
 
-# Usage
+`quickmt-zh-en` is a reasonably fast and reasonably accurate neural machine translation model for translation from `zh` into `en`.
 
-## Install `quickmt`
+`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). 
 
-```bash
-git clone https://github.com/quickmt/quickmt.git
-pip install ./quickmt/
-```
 
-## Download model
+## *UPDATED VERSION!* 
+
+This model was trained with back-translated data and has improved translation quality!
+
+* https://huggingface.co/datasets/quickmt/madlad400-en-backtranslated-zh
+* https://huggingface.co/datasets/quickmt/newscrawl2024-en-backtranslated-zh
+
+
+## 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) 
+* 200M parameter transformer 'big' with 8 encoder layers and 2 decoder layers
+* 32k separate Sentencepiece vocabs
+* Exported for fast inference to [CTranslate2](https://github.com/OpenNMT/CTranslate2) format
+* 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.
+
+
+## Usage with `quickmt`
+
+You must install the Nvidia cuda toolkit first, if you want to do GPU inference.
+
+Next, install the `quickmt` python library and download the model:
 
 ```bash
+git clone https://github.com/quickmt/quickmt.git
+pip install -e ./quickmt/
+
 quickmt-model-download quickmt/quickmt-zh-en ./quickmt-zh-en
 ```
 
-## Use model
+Finally use the model in python:
 
 ```python
 from quickmt import Translator
@@ -23,135 +85,36 @@ from quickmt import Translator
 # Auto-detects GPU, set to "cpu" to force CPU inference
 t = Translator("./quickmt-zh-en/", device="auto")
 
-# Translate - set beam size to 5 for higher quality (but slower speed)
-t(["他补充道：“我们现在有 4 个月大没有糖尿病的老鼠，但它们曾经得过该病。”"], beam_size=1)
+# Translate - set beam size to 1 for faster speed (but lower quality)
+sample_text = '埃胡德·乌尔博士（新斯科舍省哈利法克斯市达尔豪西大学医学教授，加拿大糖尿病协会临床与科学部门教授）提醒，这项研究仍处在早期阶段。'
+
+t(sample_text, beam_size=5)
+```
+
+> 'Dr. Ehud Ur (Professor of Medicine, Dalhousie University, Halifax, Nova Scotia, and Professor of Clinical and Scientific Division, Canadian Diabetes Association) cautions that the study is still at an early stage.'
 
+```python
 # Get alternative translations by sampling
 # You can pass any cTranslate2 `translate_batch` arguments
-t(["他补充道：“我们现在有 4 个月大没有糖尿病的老鼠，但它们曾经得过该病。”"], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)
+t([sample_text], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)
 ```
 
-# Model Information
+> 'Dr Elhoud (Professor of Medicine at Dalhousie University, Halifax, Nova Scotia, and professor of clinical and scientific Division of the Canadian Diabetes Association) cautions that the study is still at an early stage.'
+
+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.
 
-* Trained using [`eole`](https://github.com/eole-nlp/eole)
-* Exported for fast inference to []CTranslate2](https://github.com/OpenNMT/CTranslate2) format
-* Training data: https://huggingface.co/datasets/quickmt/quickmt-train.zh-en/tree/main
 
 ## Metrics
 
-BLEU and CHRF2 calculated with [sacrebleu](https://github.com/mjpost/sacrebleu) on the Flores200 `devtest` test set ("zho_Hans"->"eng_Latn").
-
-| Model | bleu | chrf2 |
-| ----  | ---- | ----  |
-| quickmt/quickmt-zh-en      | 28.58 | 57.46 |
-| Helsinki-NLP/opus-mt-zh-en | 23.35 | 53.60 |
-| facebook/m2m100_418M | 18.96 | 50.06 |
-| facebook/m2m100_1.2B | 24.68 | 54.68 |
-| facebook/nllb-200-distilled-600M | 26.22 | 55.17 | 
-| facebook/nllb-200-distilled-1.3B | 28.54 | 57.34 |
-| google/madlad400-3b-mt | 28.74 | 58.01 | 
-
-## Training Configuration
-
-```yaml
-## IO
-save_data: zh_en/data_spm
-overwrite: True
-seed: 1234
-report_every: 100
-valid_metrics: ["BLEU"]
-tensorboard: true
-tensorboard_log_dir: tensorboard
-
-### Vocab
-src_vocab: zh-en/src.eole.vocab
-tgt_vocab: zh-en/tgt.eole.vocab
-src_vocab_size: 20000
-tgt_vocab_size: 20000
-vocab_size_multiple: 8
-share_vocab: False
-n_sample: 0
-
-data:
-    corpus_1:
-        path_src: hf://quickmt/quickmt-train-zh-en/zh
-        path_tgt: hf://quickmt/quickmt-train-zh-en/en
-        path_sco: hf://quickmt/quickmt-train-zh-en/sco
-
-    valid:
-        path_src: zh-en/dev.zho
-        path_tgt: zh-en/dev.eng
-
-transforms: [sentencepiece, filtertoolong]
-transforms_configs:
-  sentencepiece:
-    src_subword_model: "zh-en/src.spm.model"
-    tgt_subword_model: "zh-en/tgt.spm.model"
-  filtertoolong:
-    src_seq_length: 512
-    tgt_seq_length: 512
-
-training:
-    # Run configuration
-    model_path: quickmt-zh-en
-    keep_checkpoint: 4
-    save_checkpoint_steps: 1000
-    train_steps: 200000
-    valid_steps: 1000
-    
-    # Train on a single GPU
-    world_size: 1
-    gpu_ranks: [0]
-
-    # Batching
-    batch_type: "tokens"
-    batch_size: 13312
-    valid_batch_size: 13312
-    batch_size_multiple: 8
-    accum_count: [4]
-    accum_steps: [0]
-
-    # Optimizer & Compute
-    compute_dtype: "bfloat16"
-    optim: "pagedadamw8bit"
-    learning_rate: 1.0
-    warmup_steps: 10000
-    decay_method: "noam"
-    adam_beta2: 0.998
-
-    # Data loading
-    bucket_size: 262144
-    num_workers: 4
-    prefetch_factor: 100
-
-    # 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"
-    layer_norm: standard
-    share_embeddings: false
-    share_decoder_embeddings: true
-    add_ffnbias: true
-    mlp_activation_fn: gated-silu
-    add_estimator: false
-    add_qkvbias: false
-    norm_eps: 1e-6
-    hidden_size: 1024
-    encoder:
-        layers: 8
-    decoder:
-        layers: 2
-    heads: 16
-    transformer_ff: 4096
-    embeddings:
-        word_vec_size: 1024
-        position_encoding_type: "SinusoidalInterleaved"
-```
+`bleu` and `chrf2` are calculated with [sacrebleu](https://github.com/mjpost/sacrebleu) on the [Flores200 `devtest` test set](https://huggingface.co/datasets/facebook/flores) ("zho_Hans"->"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 the flores-devtest dataset (1012 sentences) on an RTX 4070s GPU with batch size 32.
+
+
+|                                  |   bleu |   chrf2 |   comet22 |   Time (s) |
+|:---------------------------------|-------:|--------:|----------:|-----------:|
+| quickmt/quickmt-zh-en            |  29.9  |   58.42 |     86.59 |       1.22 |
+| Helsinki-NLP/opus-mt-zh-en       |  22.99 |   53.98 |     84.6  |       3.73 |
+| facebook/nllb-200-distilled-600M |  26.02 |   55.27 |     85.1  |      21.69 |
+| facebook/nllb-200-distilled-1.3B |  28.61 |   57.43 |     86.22 |      37.55 |
+| facebook/m2m100_418M             |  19.55 |   50.83 |     82.04 |      18.2  |
+| facebook/m2m100_1.2B             |  24.9  |   54.89 |     85.1  |      35.49 |
+

.ipynb_checkpoints/eole-config-checkpoint.yaml

@@ -0,0 +1,106 @@
+## IO
+save_data: data
+overwrite: True
+seed: 1234
+report_every: 100
+valid_metrics: ["BLEU"]
+tensorboard: true
+tensorboard_log_dir: tensorboard
+
+### Vocab
+src_vocab: zh.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.is-en/zh
+        path_tgt: hf://quickmt/quickmt-train.is-en/en
+        path_sco: hf://quickmt/quickmt-train.is-en/sco
+        weight: 2
+    corpus_2:
+        path_src: hf://quickmt/newscrawl2024-en-backtranslated-zh/zh
+        path_tgt: hf://quickmt/newscrawl2024-en-backtranslated-zh/en
+        path_sco: hf://quickmt/newscrawl2024-en-backtranslated-zh/sco
+        weight: 1
+    corpus_3:
+        path_src: hf://quickmt/madlad400-en-backtranslated-zh/zh
+        path_tgt: hf://quickmt/madlad400-en-backtranslated-zh/en
+        path_sco: hf://quickmt/madlad400-en-backtranslated-zh/sco
+        weight: 2
+    valid:
+        path_src: valid.zh
+        path_tgt: valid.en
+
+transforms: [sentencepiece, filtertoolong]
+transforms_configs:
+  sentencepiece:
+    src_subword_model: "zh.spm.model"
+    tgt_subword_model: "en.spm.model"
+  filtertoolong:
+    src_seq_length: 256
+    tgt_seq_length: 256
+
+training:
+    # Run configuration
+    model_path: quickmt-zh-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]
+
+    # Batching
+    batch_type: "tokens"
+    batch_size: 6000
+    valid_batch_size: 2048
+    batch_size_multiple: 8
+    accum_count: [20]
+    accum_steps: [0]
+
+    # Optimizer & Compute
+    compute_dtype: "fp16"
+    optim: "adamw"
+    #use_amp: False
+    learning_rate: 3.0
+    warmup_steps: 5000
+    decay_method: "noam"
+    adam_beta2: 0.998
+
+    # Data loading
+    bucket_size: 128000
+    num_workers: 4
+    prefetch_factor: 32
+    
+    # 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: true
+    hidden_size: 1024
+    encoder:
+        layers: 8
+    decoder:
+        layers: 2
+    heads: 8
+    transformer_ff: 4096
+    embeddings:
+        word_vec_size: 1024
+        position_encoding_type: "SinusoidalInterleaved"
+

README.md

@@ -1,12 +1,14 @@
 ---
 language:
-- zh
 - en
+- zh
 tags:
 - translation
 license: cc-by-4.0
 datasets:
 - quickmt/quickmt-train.zh-en
+- quickmt/madlad400-en-backtranslated-zh
+- quickmt/newscrawl2024-en-backtranslated-zh
 model-index:
 - name: quickmt-zh-en
   results:
@@ -21,10 +23,13 @@ model-index:
     metrics:
     - name: BLEU
       type: bleu
-      value: 29.36
+      value: 29.9
     - name: CHRF
       type: chrf
-      value: 58.10
+      value: 58.42
+    - name: COMET
+      type: comet
+      value: 86.59
 ---
 
 
@@ -32,57 +37,84 @@ model-index:
 
 `quickmt-zh-en` is a reasonably fast and reasonably accurate neural machine translation model for translation from `zh` 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 VERSION!* 
+
+This model was trained with back-translated data and has improved translation quality!
+
+* https://huggingface.co/datasets/quickmt/madlad400-en-backtranslated-zh
+* https://huggingface.co/datasets/quickmt/newscrawl2024-en-backtranslated-zh
+
+
+## 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)
+* Trained using [`eole`](https://github.com/eole-nlp/eole) 
 * 200M parameter transformer 'big' with 8 encoder layers and 2 decoder layers
-* Separate source and target Sentencepiece tokenizers
+* 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.zh-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. 
+See the `eole` model configuration in this repository for further details and the `eole-model` for the raw `eole` (pytorch) model.
 
 
 ## Usage with `quickmt`
 
-First, install `quickmt` and download the model
+You must install the Nvidia cuda toolkit first, if you want to do GPU inference.
+
+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-zh-en ./quickmt-zh-en
 ```
 
+Finally use the model in python:
+
 ```python
 from quickmt import Translator
 
 # Auto-detects GPU, set to "cpu" to force CPU inference
 t = Translator("./quickmt-zh-en/", device="auto")
 
-# Translate - set beam size to 5 for higher quality (but slower speed)
-t(["他补充道：“我们现在有 4 个月大没有糖尿病的老鼠，但它们曾经得过该病。”"], beam_size=1)
+# Translate - set beam size to 1 for faster speed (but lower quality)
+sample_text = '埃胡德·乌尔博士（新斯科舍省哈利法克斯市达尔豪西大学医学教授，加拿大糖尿病协会临床与科学部门教授）提醒，这项研究仍处在早期阶段。'
+
+t(sample_text, beam_size=5)
+```
 
+> 'Dr. Ehud Ur (Professor of Medicine, Dalhousie University, Halifax, Nova Scotia, and Professor of Clinical and Scientific Division, Canadian Diabetes Association) cautions that the study is still at an early stage.'
+
+```python
 # Get alternative translations by sampling
 # You can pass any cTranslate2 `translate_batch` arguments
-t(["他补充道：“我们现在有 4 个月大没有糖尿病的老鼠，但它们曾经得过该病。”"], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)
+t([sample_text], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)
 ```
 
-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 Elhoud (Professor of Medicine at Dalhousie University, Halifax, Nova Scotia, and professor of clinical and scientific Division of the Canadian Diabetes Association) cautions that the study is still at an early stage.'
+
+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 calculated with [sacrebleu](https://github.com/mjpost/sacrebleu) on the [Flores200 `devtest` test set](https://huggingface.co/datasets/facebook/flores) ("zho_Hans"->"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. 
+`bleu` and `chrf2` are calculated with [sacrebleu](https://github.com/mjpost/sacrebleu) on the [Flores200 `devtest` test set](https://huggingface.co/datasets/facebook/flores) ("zho_Hans"->"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 the flores-devtest dataset (1012 sentences) on an RTX 4070s GPU with batch size 32.
+
 
-| Model                            | bleu  | chrf2 | comet22  | Time (s) |
-| -------------------------------- | ----- | ----- | ----     |  ----    |
-| quickmt/quickmt-zh-en            | 29.36 | 58.10 | 0.8655   | 0.88     |
-| Helsinki-NLP/opus-mt-zh-en       | 23.35 | 53.60 | 0.8426   | 3.78     |
-| facebook/m2m100_418M             | 15.99 | 50.13 | 0.7881   | 16.61    |
-| facebook/nllb-200-distilled-600M | 26.22 | 55.18 | 0.8507   | 20.89    |
-| facebook/m2m100_1.2B             | 20.30 | 54.23 | 0.8206   | 33.12    |
-| facebook/nllb-200-distilled-1.3B | 28.56 | 57.35 | 0.8620   | 36.64    |
+|                                  |   bleu |   chrf2 |   comet22 |   Time (s) |
+|:---------------------------------|-------:|--------:|----------:|-----------:|
+| quickmt/quickmt-zh-en            |  29.9  |   58.42 |     86.59 |       1.22 |
+| Helsinki-NLP/opus-mt-zh-en       |  22.99 |   53.98 |     84.6  |       3.73 |
+| facebook/nllb-200-distilled-600M |  26.02 |   55.27 |     85.1  |      21.69 |
+| facebook/nllb-200-distilled-1.3B |  28.61 |   57.43 |     86.22 |      37.55 |
+| facebook/m2m100_418M             |  19.55 |   50.83 |     82.04 |      18.2  |
+| facebook/m2m100_1.2B             |  24.9  |   54.89 |     85.1  |      35.49 |
 
-`quickmt-zh-en` is the fastest *and* highest quality.

eole-config.yaml

@@ -1,5 +1,5 @@
 ## IO
-save_data: zh_en/data_spm
+save_data: data
 overwrite: True
 seed: 1234
 report_every: 100
@@ -8,39 +8,50 @@ tensorboard: true
 tensorboard_log_dir: tensorboard
 
 ### Vocab
-src_vocab: zh-en/src.eole.vocab
-tgt_vocab: zh-en/tgt.eole.vocab
+src_vocab: zh.eole.vocab
+tgt_vocab: en.eole.vocab
 src_vocab_size: 32000
 tgt_vocab_size: 32000
 vocab_size_multiple: 8
-share_vocab: False
+share_vocab: false
 n_sample: 0
 
 data:
     corpus_1:
-        path_src: hf://quickmt/quickmt-train-zh-en/zh
-        path_tgt: hf://quickmt/quickmt-train-zh-en/en
-        path_sco: hf://quickmt/quickmt-train-zh-en/sco
+        path_src: hf://quickmt/quickmt-train.is-en/zh
+        path_tgt: hf://quickmt/quickmt-train.is-en/en
+        path_sco: hf://quickmt/quickmt-train.is-en/sco
+        weight: 2
+    corpus_2:
+        path_src: hf://quickmt/newscrawl2024-en-backtranslated-zh/zh
+        path_tgt: hf://quickmt/newscrawl2024-en-backtranslated-zh/en
+        path_sco: hf://quickmt/newscrawl2024-en-backtranslated-zh/sco
+        weight: 1
+    corpus_3:
+        path_src: hf://quickmt/madlad400-en-backtranslated-zh/zh
+        path_tgt: hf://quickmt/madlad400-en-backtranslated-zh/en
+        path_sco: hf://quickmt/madlad400-en-backtranslated-zh/sco
+        weight: 2
     valid:
-        path_src: zh-en/dev.zho
-        path_tgt: zh-en/dev.eng
+        path_src: valid.zh
+        path_tgt: valid.en
 
 transforms: [sentencepiece, filtertoolong]
 transforms_configs:
   sentencepiece:
-    src_subword_model: "zh-en/src.spm.model"
-    tgt_subword_model: "zh-en/tgt.spm.model"
+    src_subword_model: "zh.spm.model"
+    tgt_subword_model: "en.spm.model"
   filtertoolong:
     src_seq_length: 256
     tgt_seq_length: 256
 
 training:
     # Run configuration
-    model_path: zh-en/model
+    model_path: quickmt-zh-en-eole-model
     keep_checkpoint: 4
-    save_checkpoint_steps: 2000
-    train_steps: 100000
-    valid_steps: 2000
+    train_steps: 200000
+    save_checkpoint_steps: 5000
+    valid_steps: 5000
     
     # Train on a single GPU
     world_size: 1
@@ -48,30 +59,31 @@ training:
 
     # Batching
     batch_type: "tokens"
-    batch_size: 8192
-    valid_batch_size: 8192
+    batch_size: 6000
+    valid_batch_size: 2048
     batch_size_multiple: 8
-    accum_count: [16]
+    accum_count: [20]
     accum_steps: [0]
 
     # Optimizer & Compute
-    compute_dtype: "bf16"
-    optim: "pagedadamw8bit"
-    learning_rate: 2.0
-    warmup_steps: 10000
+    compute_dtype: "fp16"
+    optim: "adamw"
+    #use_amp: False
+    learning_rate: 3.0
+    warmup_steps: 5000
     decay_method: "noam"
     adam_beta2: 0.998
 
     # Data loading
     bucket_size: 128000
     num_workers: 4
-    prefetch_factor: 100
+    prefetch_factor: 32
     
     # Hyperparams
     dropout_steps: [0]
     dropout: [0.1]
     attention_dropout: [0.1]
-    max_grad_norm: 2
+    max_grad_norm: 0
     label_smoothing: 0.1
     average_decay: 0.0001
     param_init_method: xavier_uniform
@@ -79,21 +91,16 @@ training:
 
 model:
     architecture: "transformer"
-    layer_norm: standard
     share_embeddings: false
     share_decoder_embeddings: true
-    add_ffnbias: true
-    mlp_activation_fn: gelu
-    add_estimator: false
-    add_qkvbias: false
-    norm_eps: 1e-6
     hidden_size: 1024
     encoder:
         layers: 8
     decoder:
         layers: 2
-    heads: 16
+    heads: 8
     transformer_ff: 4096
     embeddings:
         word_vec_size: 1024
         position_encoding_type: "SinusoidalInterleaved"
+

eole-model/config.json

@@ -0,0 +1,153 @@
+{
+  "n_sample": 0,
+  "share_vocab": false,
+  "report_every": 100,
+  "tgt_vocab_size": 32000,
+  "tensorboard_log_dir": "tensorboard",
+  "tensorboard_log_dir_dated": "tensorboard/Nov-28_15-33-54",
+  "valid_metrics": [
+    "BLEU"
+  ],
+  "src_vocab": "zh.eole.vocab",
+  "tensorboard": true,
+  "seed": 1234,
+  "tgt_vocab": "en.eole.vocab",
+  "vocab_size_multiple": 8,
+  "transforms": [
+    "sentencepiece",
+    "filtertoolong"
+  ],
+  "src_vocab_size": 32000,
+  "overwrite": true,
+  "save_data": "data",
+  "training": {
+    "num_workers": 0,
+    "label_smoothing": 0.1,
+    "accum_count": [
+      20
+    ],
+    "valid_steps": 5000,
+    "gpu_ranks": [
+      0
+    ],
+    "accum_steps": [
+      0
+    ],
+    "warmup_steps": 5000,
+    "world_size": 1,
+    "batch_size_multiple": 8,
+    "optim": "adamw",
+    "normalization": "tokens",
+    "max_grad_norm": 0.0,
+    "bucket_size": 128000,
+    "dropout": [
+      0.1
+    ],
+    "adam_beta2": 0.998,
+    "model_path": "quickmt-zh-en-eole-model",
+    "batch_size": 6000,
+    "batch_type": "tokens",
+    "compute_dtype": "torch.float16",
+    "save_checkpoint_steps": 5000,
+    "keep_checkpoint": 4,
+    "learning_rate": 3.0,
+    "prefetch_factor": 32,
+    "dropout_steps": [
+      0
+    ],
+    "train_steps": 200000,
+    "decay_method": "noam",
+    "average_decay": 0.0001,
+    "valid_batch_size": 2048,
+    "param_init_method": "xavier_uniform",
+    "attention_dropout": [
+      0.1
+    ]
+  },
+  "transforms_configs": {
+    "sentencepiece": {
+      "src_subword_model": "${MODEL_PATH}/zh.spm.model",
+      "tgt_subword_model": "${MODEL_PATH}/en.spm.model"
+    },
+    "filtertoolong": {
+      "src_seq_length": 256,
+      "tgt_seq_length": 256
+    }
+  },
+  "data": {
+    "corpus_1": {
+      "weight": 2,
+      "transforms": [
+        "sentencepiece",
+        "filtertoolong"
+      ],
+      "path_align": null,
+      "path_src": "train.zh",
+      "path_tgt": "train.en"
+    },
+    "corpus_2": {
+      "weight": 1,
+      "transforms": [
+        "sentencepiece",
+        "filtertoolong"
+      ],
+      "path_align": null,
+      "path_src": "/home/mark/mt/data/newscrawl.backtrans.zh",
+      "path_tgt": "/home/mark/mt/data/newscrawl.2024.en"
+    },
+    "corpus_3": {
+      "weight": 2,
+      "transforms": [
+        "sentencepiece",
+        "filtertoolong"
+      ],
+      "path_align": null,
+      "path_src": "/home/mark/mt/data/madlad.backtrans.zh",
+      "path_tgt": "/home/mark/mt/data/madlad.en"
+    },
+    "valid": {
+      "path_src": "valid.zh",
+      "transforms": [
+        "sentencepiece",
+        "filtertoolong"
+      ],
+      "path_tgt": "valid.en",
+      "path_align": null
+    }
+  },
+  "model": {
+    "hidden_size": 1024,
+    "position_encoding_type": "SinusoidalInterleaved",
+    "share_embeddings": false,
+    "architecture": "transformer",
+    "heads": 8,
+    "share_decoder_embeddings": true,
+    "transformer_ff": 4096,
+    "decoder": {
+      "hidden_size": 1024,
+      "layers": 2,
+      "position_encoding_type": "SinusoidalInterleaved",
+      "tgt_word_vec_size": 1024,
+      "n_positions": null,
+      "heads": 8,
+      "decoder_type": "transformer",
+      "transformer_ff": 4096
+    },
+    "embeddings": {
+      "src_word_vec_size": 1024,
+      "word_vec_size": 1024,
+      "position_encoding_type": "SinusoidalInterleaved",
+      "tgt_word_vec_size": 1024
+    },
+    "encoder": {
+      "hidden_size": 1024,
+      "encoder_type": "transformer",
+      "src_word_vec_size": 1024,
+      "layers": 8,
+      "position_encoding_type": "SinusoidalInterleaved",
+      "n_positions": null,
+      "heads": 8,
+      "transformer_ff": 4096
+    }
+  }
+}

eole-model/en.spm.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c23dc1aa66b7b98263da867137a7eb41d5e4573984fb100c0b295f3010381823
+size 792100

eole-model/model.00.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce0b3dfe2ef4c9f6b93f969f27f5d5cf38432e4a7bcd144577c8583209bb701a
+size 840314816

eole-model/zh.spm.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56159dc3a8607805f0e3f2cf4c91f37ee221e91fb824ebb362622d22185872cb
+size 720056

model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:494518d6282fcd01f850ab4ab096e6a5c937aa834290a62e5efd435275828c9d
-size 409972810
+oid sha256:88ef37879afce2d5f0bdf4c53073aab30967f178f0a0fa2eed7c98160270b06a
+size 409915789

src.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:23d03d562fc3f8fe57e497dac0ece4827c254675a80c103fc4bb4040638ceb67
-size 733978
+oid sha256:56159dc3a8607805f0e3f2cf4c91f37ee221e91fb824ebb362622d22185872cb
+size 720056

tgt.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c373f1d78753313b0dbc337058bf8450e1fdd6fe662a49e0941affce44ec14c5
-size 800955
+oid sha256:c23dc1aa66b7b98263da867137a7eb41d5e4573984fb100c0b295f3010381823
+size 792100