Upload folder using huggingface_hub

.ipynb_checkpoints/README-checkpoint.md

@@ -1,12 +1,14 @@
 ---
 language:
-- en
 - fa
+- en
 tags:
 - translation
 license: cc-by-4.0
 datasets:
 - quickmt/quickmt-train.fa-en
+- quickmt/madlad400-en-backtranslated-fa
+- quickmt/newscrawl2024-en-backtranslated-fa
 model-index:
 - name: quickmt-fa-en
   results:
@@ -21,13 +23,13 @@ model-index:
     metrics:
     - name: BLEU
       type: bleu
-      value: 37.57
+      value: 38.99
     - name: CHRF
       type: chrf
-      value: 63.37
+      value: 64.55
     - name: COMET
       type: comet
-      value: 87.76
+      value: 88.14
 ---
 
 
@@ -35,17 +37,30 @@ model-index:
 
 `quickmt-fa-en` is a reasonably fast and reasonably accurate neural machine translation model for translation from `fa` 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!
+
+
+## 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
-* 50k joint Sentencepiece vocabulary
-* Expested for fast inference to [CTranslate2](https://github.com/OpenNMT/CTranslate2) format
-* Training data: https://huggingface.co/datasets/quickmt/quickmt-train.fa-en/tree/main
+* 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
+* 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.
@@ -54,7 +69,7 @@ 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-fa-en ./quickmt-fa-en
 ```
@@ -62,37 +77,43 @@ quickmt-model-download quickmt/quickmt-fa-en ./quickmt-fa-en
 Finally use the model in python:
 
 ```python
-from quickmt impest Translator
+from quickmt import Translator
 
 # Auto-detects GPU, set to "cpu" to force CPU inference
-t = Translator("./quickmt-fa-en/", device="auto")
+mt = Translator("./quickmt-ar-en/", device="auto")
 
 # Translate - set beam size to 1 for faster speed (but lower quality)
 sample_text = 'دکتر ایهود اور، استاد پزشکی دانشگاه دالهاوزی در هلیفکس، نوااسکوشیا و رئیس بخش کلینیکی و علمی انجمن دیابت کانادا هشدار داد که این تحقیق هنوز در روزهای آغازین خود می\u200cباشد.'
 
-t(sample_text, beam_size=5)
+mt(sample_text, beam_size=5)
 ```
 
-> 'Dr. Ehudover, a professor of medicine at Dalhousie University in Halifax, Nova Scotia and head of the clinical and scientific division of the Canadian Diabetes Association, warned that the study was still in its early days.'
+> 'Dr. Ehud Orr, a professor of medicine at Dalhousie University in Halifax, Nova Scotia, and head of the Canadian Diabetes Association’s clinical and scientific department, warned that the research is still in its early days.'
 
 ```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)
+mt([sample_text], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)
 ```
 
-> 'Dr. Ehudver, professor of medicine at Dalhousie University in Halifax, Nova Scotia and head of the Clinical and Scientific Section of the Canadian Diabetes Society, cautioned the new study was still very early.'
+> 'Dr. Ehud Orr, medical professor of Dalhousie University in Halifax, Nova Scotia and head of the Clinical and Scientific Section of the Canadian Diabetes Association warned that the research is still in its early days.'
+
+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.
 
-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`.
 
 ## 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) ("pes_Arab"->"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 (faster speed is possible using a larger 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-fa-en                 |  38.99 |   64.55 |     88.14 |       1.1  |
+| facebook/nllb-200-distilled-600M      |  34.8  |   60.86 |     86.49 |      21.13 |
+| facebook/nllb-200-distilled-1.3B      |  37.91 |   63.39 |     87.82 |      36.86 |
+| facebook/m2m100_418M                  |  27.2  |   55.82 |     82.9  |      18.23 |
+| facebook/m2m100_1.2B                  |  29.13 |   56.4  |     83.5  |      34.8  |
+| tencent/HY-MT1.5-1.8B                 |  20.87 |   55.02 |     86.22 |      10.0  |
+| tencent/HY-MT1.5-7B-FP8               |  28.16 |   59.49 |     88.07 |      36.0  |
+| CohereLabs/aya-expanse-8b (bnb quant) |  35.29 |   62.46 |     88.43 |      77.37 |
 
-|                                  |   bleu |   chrf2 |   comet22 |   Time (s) |
-|:---------------------------------|-------:|--------:|----------:|-----------:|
-| quickmt-fa-en                    |  37.57 |   63.37 |     87.76 |       1.16 |
-| facebook/nllb-200-distilled-600M |  34.79 |   60.86 |     86.49 |      21.17 |
-| facebook/nllb-200-distilled-1.3B |  37.91 |   63.39 |     87.82 |      36.9  |
-| facebook/m2m100_418M             |  27.2  |   55.82 |     82.9  |      18.24 |
-| facebook/m2m100_1.2B             |  29.12 |   56.39 |     83.5  |      35.14 |

README.md

@@ -1,12 +1,14 @@
 ---
 language:
-- en
 - fa
+- en
 tags:
 - translation
 license: cc-by-4.0
 datasets:
 - quickmt/quickmt-train.fa-en
+- quickmt/madlad400-en-backtranslated-fa
+- quickmt/newscrawl2024-en-backtranslated-fa
 model-index:
 - name: quickmt-fa-en
   results:
@@ -21,13 +23,13 @@ model-index:
     metrics:
     - name: BLEU
       type: bleu
-      value: 37.57
+      value: 38.99
     - name: CHRF
       type: chrf
-      value: 63.37
+      value: 64.55
     - name: COMET
       type: comet
-      value: 87.76
+      value: 88.14
 ---
 
 
@@ -35,17 +37,30 @@ model-index:
 
 `quickmt-fa-en` is a reasonably fast and reasonably accurate neural machine translation model for translation from `fa` 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!
+
+
+## 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
-* 50k joint Sentencepiece vocabulary
-* Expested for fast inference to [CTranslate2](https://github.com/OpenNMT/CTranslate2) format
-* Training data: https://huggingface.co/datasets/quickmt/quickmt-train.fa-en/tree/main
+* 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
+* 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.
@@ -54,7 +69,7 @@ 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-fa-en ./quickmt-fa-en
 ```
@@ -65,34 +80,40 @@ Finally use the model in python:
 from quickmt import Translator
 
 # Auto-detects GPU, set to "cpu" to force CPU inference
-t = Translator("./quickmt-fa-en/", device="auto")
+mt = Translator("./quickmt-ar-en/", device="auto")
 
 # Translate - set beam size to 1 for faster speed (but lower quality)
 sample_text = 'دکتر ایهود اور، استاد پزشکی دانشگاه دالهاوزی در هلیفکس، نوااسکوشیا و رئیس بخش کلینیکی و علمی انجمن دیابت کانادا هشدار داد که این تحقیق هنوز در روزهای آغازین خود می\u200cباشد.'
 
-t(sample_text, beam_size=5)
+mt(sample_text, beam_size=5)
 ```
 
-> 'Dr. Ehudover, a professor of medicine at Dalhousie University in Halifax, Nova Scotia and head of the clinical and scientific division of the Canadian Diabetes Association, warned that the study was still in its early days.'
+> 'Dr. Ehud Orr, a professor of medicine at Dalhousie University in Halifax, Nova Scotia, and head of the Canadian Diabetes Association’s clinical and scientific department, warned that the research is still in its early days.'
 
 ```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)
+mt([sample_text], sampling_temperature=1.2, beam_size=1, sampling_topk=50, sampling_topp=0.9)
 ```
 
-> 'Dr. Ehudver, professor of medicine at Dalhousie University in Halifax, Nova Scotia and head of the Clinical and Scientific Section of the Canadian Diabetes Society, cautioned the new study was still very early.'
+> 'Dr. Ehud Orr, medical professor of Dalhousie University in Halifax, Nova Scotia and head of the Clinical and Scientific Section of the Canadian Diabetes Association warned that the research is still in its early days.'
+
+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.
 
-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`.
 
 ## 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) ("pes_Arab"->"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 (faster speed is possible using a larger 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-fa-en                 |  38.99 |   64.55 |     88.14 |       1.1  |
+| facebook/nllb-200-distilled-600M      |  34.8  |   60.86 |     86.49 |      21.13 |
+| facebook/nllb-200-distilled-1.3B      |  37.91 |   63.39 |     87.82 |      36.86 |
+| facebook/m2m100_418M                  |  27.2  |   55.82 |     82.9  |      18.23 |
+| facebook/m2m100_1.2B                  |  29.13 |   56.4  |     83.5  |      34.8  |
+| tencent/HY-MT1.5-1.8B                 |  20.87 |   55.02 |     86.22 |      10.0  |
+| tencent/HY-MT1.5-7B-FP8               |  28.16 |   59.49 |     88.07 |      36.0  |
+| CohereLabs/aya-expanse-8b (bnb quant) |  35.29 |   62.46 |     88.43 |      77.37 |
 
-|                                  |   bleu |   chrf2 |   comet22 |   Time (s) |
-|:---------------------------------|-------:|--------:|----------:|-----------:|
-| quickmt-fa-en                    |  37.57 |   63.37 |     87.76 |       1.16 |
-| facebook/nllb-200-distilled-600M |  34.79 |   60.86 |     86.49 |      21.17 |
-| facebook/nllb-200-distilled-1.3B |  37.91 |   63.39 |     87.82 |      36.9  |
-| facebook/m2m100_418M             |  27.2  |   55.82 |     82.9  |      18.24 |
-| facebook/m2m100_1.2B             |  29.12 |   56.39 |     83.5  |      35.14 |

eole-config.yaml

@@ -8,30 +8,36 @@ tensorboard: true
 tensorboard_log_dir: tensorboard
 
 ### Vocab
-src_vocab: fa.eole.vocab
-tgt_vocab: en.eole.vocab
-src_vocab_size: 20000
-tgt_vocab_size: 20000
+src_vocab: faen/fa.eole.vocab
+tgt_vocab: faen/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.id-en/id
-        # path_tgt: hf://quickmt/quickmt-train.id-en/en
-        # path_sco: hf://quickmt/quickmt-train.id-en/sco
-        path_src: train.fa
-        path_tgt: train.en
+        path_src: faen/train.cleaned.filtered.fa
+        path_tgt: faen/train.cleaned.filtered.en
+        weight: 2
+    corpus_2:
+        path_src: /home/mark/mt/data/newscrawl.backtrans.fa
+        path_tgt: /home/mark/mt/data/newscrawl.2024.en
+        weight: 1
+    corpus_3:
+        path_src: /home/mark/mt/data/madlad.backtrans.fa
+        path_tgt: /home/mark/mt/data/madlad.en
+        weight: 2
     valid:
-        path_src: dev.fa
-        path_tgt: dev.en
+        path_src: faen/dev.fa
+        path_tgt: faen/dev.en
 
 transforms: [sentencepiece, filtertoolong]
 transforms_configs:
   sentencepiece:
-    src_subword_model: "fa.spm.model"
-    tgt_subword_model: "en.spm.model"
+    src_subword_model: "faen/fa.spm.model"
+    tgt_subword_model: "faen/en.spm.model"
   filtertoolong:
     src_seq_length: 256
     tgt_seq_length: 256
@@ -39,9 +45,8 @@ transforms_configs:
 training:
     # Run configuration
     model_path: quickmt-fa-en-eole-model
-    #train_from: model
     keep_checkpoint: 4
-    train_steps: 100000
+    train_steps: 200000
     save_checkpoint_steps: 5000
     valid_steps: 5000
     
@@ -49,27 +54,28 @@ training:
     world_size: 1
     gpu_ranks: [0]
 
-    # Batching 10240
+    # Batching 120,000 tokens
+    # For RTX 5090, 15000 batch size, accum_count 8
     batch_type: "tokens"
-    batch_size: 8000
-    valid_batch_size: 4096
+    batch_size: 6000
+    valid_batch_size: 2048
     batch_size_multiple: 8
-    accum_count: [10]
+    accum_count: [20]
     accum_steps: [0]
 
     # Optimizer & Compute
     compute_dtype: "fp16"
     optim: "adamw"
-    #use_amp: False
-    learning_rate: 2.0
-    warmup_steps: 4000
+    #use_amp: True
+    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: 128
     
     # Hyperparams
     dropout_steps: [0]
@@ -85,14 +91,20 @@ 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

@@ -1,68 +1,68 @@
 {
-  "share_vocab": false,
-  "vocab_size_multiple": 8,
-  "src_vocab_size": 20000,
-  "seed": 1234,
-  "valid_metrics": [
-    "BLEU"
+  "transforms": [
+    "sentencepiece",
+    "filtertoolong"
   ],
+  "share_vocab": false,
   "save_data": "data",
-  "tensorboard": true,
-  "tgt_vocab": "en.eole.vocab",
   "tensorboard_log_dir": "tensorboard",
-  "tensorboard_log_dir_dated": "tensorboard/Jun-03_06-03-40",
   "overwrite": true,
-  "tgt_vocab_size": 20000,
-  "src_vocab": "fa.eole.vocab",
-  "n_sample": 0,
-  "report_every": 100,
-  "transforms": [
-    "sentencepiece",
-    "filtertoolong"
+  "valid_metrics": [
+    "BLEU"
   ],
+  "report_every": 100,
+  "tgt_vocab": "faen/en.eole.vocab",
+  "vocab_size_multiple": 8,
+  "n_sample": 0,
+  "seed": 1234,
+  "tensorboard": true,
+  "tensorboard_log_dir_dated": "tensorboard/Jan-02_10-23-54",
+  "src_vocab_size": 32000,
+  "src_vocab": "faen/fa.eole.vocab",
+  "tgt_vocab_size": 32000,
   "training": {
-    "train_steps": 100000,
     "dropout_steps": [
       0
     ],
-    "batch_size": 8000,
-    "attention_dropout": [
-      0.1
-    ],
-    "accum_count": [
-      10
-    ],
-    "prefetch_factor": 32,
-    "valid_steps": 5000,
-    "adam_beta2": 0.998,
     "world_size": 1,
+    "warmup_steps": 5000,
+    "num_workers": 0,
+    "batch_size_multiple": 8,
+    "compute_dtype": "torch.float16",
+    "param_init_method": "xavier_uniform",
+    "normalization": "tokens",
     "accum_steps": [
       0
     ],
-    "model_path": "quickmt-fa-en-eole-model",
-    "label_smoothing": 0.1,
-    "keep_checkpoint": 4,
-    "gpu_ranks": [
-      0
-    ],
-    "batch_size_multiple": 8,
-    "warmup_steps": 4000,
     "decay_method": "noam",
-    "max_grad_norm": 0.0,
+    "model_path": "quickmt-fa-en-eole-model",
+    "prefetch_factor": 128,
     "batch_type": "tokens",
+    "valid_batch_size": 2048,
     "save_checkpoint_steps": 5000,
-    "param_init_method": "xavier_uniform",
-    "normalization": "tokens",
-    "learning_rate": 2.0,
-    "optim": "adamw",
-    "compute_dtype": "torch.float16",
-    "bucket_size": 128000,
+    "train_steps": 200000,
     "dropout": [
       0.1
     ],
+    "attention_dropout": [
+      0.1
+    ],
+    "batch_size": 6000,
+    "label_smoothing": 0.1,
     "average_decay": 0.0001,
-    "num_workers": 0,
-    "valid_batch_size": 4096
+    "learning_rate": 3.0,
+    "max_grad_norm": 0.0,
+    "accum_count": [
+      20
+    ],
+    "gpu_ranks": [
+      0
+    ],
+    "keep_checkpoint": 4,
+    "bucket_size": 256000,
+    "optim": "adamw",
+    "valid_steps": 5000,
+    "adam_beta2": 0.998
   },
   "data": {
     "corpus_1": {
@@ -70,63 +70,97 @@
         "sentencepiece",
         "filtertoolong"
       ],
-      "path_src": "train.fa",
+      "weight": 2,
       "path_align": null,
-      "path_tgt": "train.en"
+      "path_tgt": "faen/train.cleaned.filtered.en",
+      "path_src": "faen/train.cleaned.filtered.fa"
     },
-    "valid": {
+    "corpus_2": {
       "transforms": [
         "sentencepiece",
         "filtertoolong"
       ],
-      "path_src": "dev.fa",
+      "weight": 1,
       "path_align": null,
-      "path_tgt": "dev.en"
-    }
-  },
-  "transforms_configs": {
-    "sentencepiece": {
-      "src_subword_model": "${MODEL_PATH}/fa.spm.model",
-      "tgt_subword_model": "${MODEL_PATH}/en.spm.model"
+      "path_tgt": "/home/mark/mt/data/newscrawl.2024.en",
+      "path_src": "/home/mark/mt/data/newscrawl.backtrans.fa"
     },
-    "filtertoolong": {
-      "src_seq_length": 256,
-      "tgt_seq_length": 256
+    "corpus_3": {
+      "transforms": [
+        "sentencepiece",
+        "filtertoolong"
+      ],
+      "weight": 2,
+      "path_align": null,
+      "path_tgt": "/home/mark/mt/data/madlad.en",
+      "path_src": "/home/mark/mt/data/madlad.backtrans.fa"
+    },
+    "valid": {
+      "transforms": [
+        "sentencepiece",
+        "filtertoolong"
+      ],
+      "path_tgt": "faen/dev.en",
+      "path_src": "faen/dev.fa",
+      "path_align": null
     }
   },
   "model": {
+    "heads": 16,
     "share_embeddings": false,
-    "hidden_size": 1024,
-    "architecture": "transformer",
-    "share_decoder_embeddings": false,
-    "heads": 8,
-    "transformer_ff": 4096,
+    "add_estimator": false,
+    "layer_norm": "standard",
     "position_encoding_type": "SinusoidalInterleaved",
-    "decoder": {
-      "decoder_type": "transformer",
-      "hidden_size": 1024,
-      "n_positions": null,
-      "transformer_ff": 4096,
-      "heads": 8,
+    "add_qkvbias": false,
+    "transformer_ff": 4096,
+    "share_decoder_embeddings": false,
+    "hidden_size": 768,
+    "architecture": "transformer",
+    "add_ffnbias": true,
+    "mlp_activation_fn": "gelu",
+    "embeddings": {
+      "word_vec_size": 768,
+      "src_word_vec_size": 768,
       "position_encoding_type": "SinusoidalInterleaved",
-      "tgt_word_vec_size": 1024,
-      "layers": 2
+      "tgt_word_vec_size": 768
     },
     "encoder": {
-      "hidden_size": 1024,
+      "heads": 16,
       "encoder_type": "transformer",
+      "layer_norm": "standard",
+      "position_encoding_type": "SinusoidalInterleaved",
+      "add_qkvbias": false,
       "n_positions": null,
-      "src_word_vec_size": 1024,
-      "heads": 8,
       "transformer_ff": 4096,
+      "src_word_vec_size": 768,
+      "layers": 12,
+      "add_ffnbias": true,
+      "hidden_size": 768,
+      "mlp_activation_fn": "gelu"
+    },
+    "decoder": {
+      "heads": 16,
+      "layer_norm": "standard",
       "position_encoding_type": "SinusoidalInterleaved",
-      "layers": 8
+      "add_qkvbias": false,
+      "decoder_type": "transformer",
+      "n_positions": null,
+      "transformer_ff": 4096,
+      "tgt_word_vec_size": 768,
+      "layers": 2,
+      "add_ffnbias": true,
+      "hidden_size": 768,
+      "mlp_activation_fn": "gelu"
+    }
+  },
+  "transforms_configs": {
+    "sentencepiece": {
+      "tgt_subword_model": "${MODEL_PATH}/en.spm.model",
+      "src_subword_model": "${MODEL_PATH}/fa.spm.model"
     },
-    "embeddings": {
-      "word_vec_size": 1024,
-      "src_word_vec_size": 1024,
-      "tgt_word_vec_size": 1024,
-      "position_encoding_type": "SinusoidalInterleaved"
+    "filtertoolong": {
+      "src_seq_length": 256,
+      "tgt_seq_length": 256
     }
   }
 }

eole-model/en.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8595a973d0f43dd25225a0eb3411ffbdf8ae4736e41d800995b1d25ac2c6019a
-size 588878
+oid sha256:53931998683cf961ec8a9d3327deb4021d5bc12be2f2650758bb9a35dd828599
+size 805200

eole-model/fa.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6aedb8783f67ae6310ebe5f9c0798f6fcdf7bfcff877d548d039d7bc53ca6548
-size 640837
+oid sha256:a0b876dcd23cce26a8b2dd35f3de1bea3dada039d27753498f5388960a97aa3f
+size 889048

eole-model/model.00.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:efd987b8011fdbe578da4e97889cb0815e6f1f34de81fdeb56ae26f2db7e6849
-size 823882912
+oid sha256:7acbcafc1d1cc9c6a51f34719614d2346ec80019d637bd23dbee20fbaed5668c
+size 829569112

model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9317a7e6b5c80b63a6827571c5a3b9fffde72d5af6aad96175d51459be94844b
-size 401699775
+oid sha256:091265c533b556b8c9d111519dd8142ff49e901256e35613876ce92679c91487
+size 407101843

src.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6aedb8783f67ae6310ebe5f9c0798f6fcdf7bfcff877d548d039d7bc53ca6548
-size 640837
+oid sha256:a0b876dcd23cce26a8b2dd35f3de1bea3dada039d27753498f5388960a97aa3f
+size 889048

tgt.spm.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8595a973d0f43dd25225a0eb3411ffbdf8ae4736e41d800995b1d25ac2c6019a
-size 588878
+oid sha256:53931998683cf961ec8a9d3327deb4021d5bc12be2f2650758bb9a35dd828599
+size 805200