base on State-of-the-art LLM-based translation models. <p align="center"> <img alt="ALMA" src="./figures/alma_logo.png" width="500" height="203"> </p> <div align="center"> # ALMA: Advanced Language Model-based translator </div> <p align="center"> <a href="LICENSE" alt="MIT License"><img src="https://img.shields.io/badge/license-MIT-FAD689.svg" /></a> <a href="https://arxiv.org/abs/2309.11674" alt="ALMA paper"><img src="https://img.shields.io/badge/ALMA-Paper-D9AB42" /></a> <a href="https://arxiv.org/abs/2401.08417" alt="ALMA-R paper"><img src="https://img.shields.io/badge/ALMA--R-Paper-F6C555" /></a> <a href="https://arxiv.org/pdf/2410.03115" alt="X-ALMA paper"><img src="https://img.shields.io/badge/X--ALMA-Paper-F3B425" /></a> <!-- <a href="https://notes.aimodels.fyi/alma-a-new-training-method-that-boosts-translation-performance-for-large-language-models/"><img alt="Summary Link" src="https://img.shields.io/badge/summary-link-F6C555" /></a> --> <a href="https://www.clsp.jhu.edu/" alt="jhu"><img src="https://img.shields.io/badge/Johns_Hopkins_University-BEC23F" /></a> <a href="https://www.microsoft.com/en-us/research/" alt="MSlogo"><img src="https://img.shields.io/badge/Microsoft-B1B479?logo=microsoft" /></a> <a href="https://twitter.com/fe1ixxu"> <img src="https://img.shields.io/twitter/follow/haoranxu?style=social&logo=twitter" alt="follow on Twitter"></a> </p> ALMA has three generations: ALMA (1st), ALMA-R (2nd), and **X-ALMA(3rd NEW!)**. [**ALMA**](https://arxiv.org/abs/2309.11674) (**A**dvanced **L**anguage **M**odel-based Tr**A**nslator) is a many-to-many LLM-based translation model, which adopts a new translation model paradigm: it begins with fine-tuning on monolingual data and is further optimized using high-quality parallel data. This two-step fine-tuning process ensures strong translation performance. **[ALMA-R](https://arxiv.org/pdf/2401.08417v2.pdf)** builds upon ALMA models, with further LoRA fine-tuning with our proposed **Contrastive Preference Optimization (CPO)** as opposed to the Supervised Fine-tuning used in ALMA. CPO fine-tuning requires our [triplet preference data](https://huggingface.co/datasets/haoranxu/ALMA-R-Preference) for preference learning. ALMA-R now can matches or even exceeds GPT-4 or WMT winners! **[X-ALMA](https://arxiv.org/pdf/2410.03115) (NEW!) extends ALMA(-R) from 6 languages to 50 languages and ensures top-tier performance across 50 diverse languages, regardless of their resource levels. This is achieved by plug-and-play language-specific module architecture and a carefully designed 5-step training recipe with novel *Adaptive-Rejection Preference Optimization* methods.** *Old ALMA Repo:* - The original **ALMA** repository can be found [here](https://github.com/fe1ixxu/ALMA/tree/a3cc7877752779346312bb07798172eadc83d692). - The original **ALMA-R** repository can be found [here](https://github.com/fe1ixxu/ALMA/tree/ac120eb44c609ad9a386d617172d40432c2c0df6). # News 🌟 ⭐ Jan. 22 2025 **X-ALMA** has been accepted at **ICLR 2025**! ⭐ Oct. 6 2024 **X-ALMA** is out! Please find the [paper here](https://arxiv.org/pdf/2410.03115) and [models & datasets here](https://huggingface.co/collections/haoranxu/x-alma-66fde464ef90be465920abaa). ⭐ Jun. 20 2024 We want to give a shout out to [SimPO](https://arxiv.org/pdf/2405.14734), which shares a similar reference-free preference learning framework with CPO but in a more stable manner due to its special length normalization and target reward margin. The most exciting thing is that CPO and SimPO can potentially be used together! Learn more about [CPO-SimPO](https://github.com/fe1ixxu/CPO_SIMPO)! ⭐ May.1 CPO paper has been accepted at **ICML 2024**! ⭐ Mar.22 2024 CPO method now is merged at [huggingface trl](https://github.com/huggingface/trl)! See details [here](https://github.com/huggingface/trl/pull/1382). ⭐ Jan.16 2024 **ALMA-R** is released! Please check more details with our new paper: [Contrastive Preference Optimization: Pushing the Boundaries of LLM Performance in Machine Translation](https://arxiv.org/abs/2401.08417). ⭐ Jan.16 2024 The ALMA paper: [A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models](https://arxiv.org/abs/2309.11674) has been accepted at **ICLR 2024**! Check out more details [here](https://openreview.net/forum?id=farT6XXntP)! # Contents 📄 - [Download ALMA Models and Dataset](#download-alma-models-and-dataset-) - [A Quick Start](#a-quick-start) - [Environment Setup](#environment-setup-) - [Evaluation](#evaluation-) - [Training](#training-) - [FAQs](#faqs-) :star: Supports :star: - AMD and Nvidia Cards - Data Parallel Evaluation - Also support LLaMA-1, LLaMA-2, OPT, Faclon, BLOOM, MPT - LoRA Fine-tuning - Monolingual data fine-tuning, parallel data fine-tuning <p align="center"> <img src="./figures/xalma.png" width="700" height="300"> </p> # Download ALMA Models and Dataset 🚀 We release seven translation models for ALMA series: Model checkpoints are released at huggingface: | Models | Base Model Link | LoRA Link | |:-------------:|:---------------:|:---------:| | ALMA-7B (1st gen) | [haoranxu/ALMA-7B](https://huggingface.co/haoranxu/ALMA-7B) | - | | ALMA-7B-LoRA (1st gen) | [haoranxu/ALMA-7B-Pretrain](https://huggingface.co/haoranxu/ALMA-7B-Pretrain) | [haoranxu/ALMA-7B-Pretrain-LoRA](https://huggingface.co/haoranxu/ALMA-7B-Pretrain-LoRA) | | ALMA-7B-R (2nd gen) | [haoranxu/ALMA-7B-R (LoRA merged)](https://huggingface.co/haoranxu/ALMA-7B-R) | - | | ALMA-13B-LoRA (1st gen) | [haoranxu/ALMA-13B](https://huggingface.co/haoranxu/ALMA-13B) | - | | ALMA-13B-LoRA | [haoranxu/ALMA-13B-Pretrain](https://huggingface.co/haoranxu/ALMA-13B-Pretrain) | [haoranxu/ALMA-13B-Pretrain-LoRA](https://huggingface.co/haoranxu/ALMA-13B-Pretrain-LoRA) | | ALMA-13B-R (2nd gen) | [haoranxu/ALMA-13B-R (LoRA merged)](https://huggingface.co/haoranxu/ALMA-13B-R) | - | | **X-ALMA (NEW, 3rd gen)** | [X-ALMA Models](https://huggingface.co/collections/haoranxu/x-alma-66fde464ef90be465920abaa) | - | **Note that `ALMA-7B-Pretrain` and `ALMA-13B-Pretrain` are NOT translation models. They only experience stage 1 monolingual fine-tuning (20B tokens for the 7B model and 12B tokens for the 13B model), and should be utilized in conjunction with their LoRA models.** *We have also provided the WMT'22 and WMT'23 translation outputs from ALMA-13B-LoRA and ALMA-13B-R in the `outputs` directory. These outputs also includes our outputs of baselines and can be directly accessed and used for subsequent evaluations.* Datasets used by ALMA and ALMA-R are also released at huggingface now (NEW!) | Datasets | Train / Validation| Test | |:-------------:|:---------------:|:---------:| | ALMA Human-Written Parallel Data | [Parallel train and validation](https://huggingface.co/datasets/haoranxu/ALMA-Human-Parallel) | [WMT'22](https://huggingface.co/datasets/haoranxu/WMT22-Test) | | ALMA-R Triplet Preference Data | [Triplet Preference Data](https://huggingface.co/datasets/haoranxu/ALMA-R-Preference) | [WMT'22](https://huggingface.co/datasets/haoranxu/WMT22-Test) and [WMT'23](https://huggingface.co/datasets/haoranxu/WMT23-Test) | | **X-ALMA Data** | 50-language [parallel data](https://huggingface.co/datasets/haoranxu/X-ALMA-Parallel-Data) and [preference data](https://huggingface.co/datasets/haoranxu/X-ALMA-Preference) | [WMT'23](https://huggingface.co/datasets/haoranxu/WMT23-Test) and [FLORES-200](https://huggingface.co/datasets/haoranxu/X-ALMA-Parallel-Data) | # A Quick Start X-ALMA is designed with a plug-and-play architecture, consisting of two components: a base model and language-specific modules, with each module shared across different language groups. There are three ways to load X-ALMA for translation. An example of translating "我爱机器翻译。" into English (X-ALMA should also able to do multilingual open-ended QA). **The first way**: loading the merged model where the language-specific module has been merged into the base model (Recommended): ``` import torch from transformers import AutoModelForCausalLM from transformers import AutoTokenizer from peft import PeftModel GROUP2LANG = { 1: ["da", "nl", "de", "is", "no", "sv", "af"], 2: ["ca", "ro", "gl", "it", "pt", "es"], 3: ["bg", "mk", "sr", "uk", "ru"], 4: ["id", "ms", "th", "vi", "mg", "fr"], 5: ["hu", "el", "cs", "pl", "lt", "lv"], 6: ["ka", "zh", "ja", "ko", "fi", "et"], 7: ["gu", "hi", "mr", "ne", "ur"], 8: ["az", "kk", "ky", "tr", "uz", "ar", "he", "fa"], } LANG2GROUP = {lang: str(group) for group, langs in GROUP2LANG.items() for lang in langs} group_id = LANG2GROUP["zh"] model = AutoModelForCausalLM.from_pretrained(f"haoranxu/X-ALMA-13B-Group{group_id}", torch_dtype=torch.float16, device_map="auto") tokenizer = AutoTokenizer.from_pretrained(f"haoranxu/X-ALMA-13B-Group{group_id}", padding_side='left') # Add the source sentence into the prompt template prompt="Translate this from Chinese to English:\nChinese: 我爱机器翻译。\nEnglish:" # X-ALMA needs chat template but ALMA and ALMA-R don't need it. chat_style_prompt = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template(chat_style_prompt, tokenize=False, add_generation_prompt=True) input_ids = tokenizer(prompt, return_tensors="pt", padding=True, max_length=40, truncation=True).input_ids.cuda() # Translation with torch.no_grad(): generated_ids = model.generate(input_ids=input_ids, num_beams=5, max_new_tokens=20, do_sample=True, temperature=0.6, top_p=0.9) outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=True) print(outputs) ``` **The second way**: loading the base model and language-specific module (Recommended): ``` model = AutoModelForCausalLM.from_pretrained("haoranxu/X-ALMA-13B-Pretrain", torch_dtype=torch.float16, device_map="auto") model = PeftModel.from_pretrained(model, f"haoranxu/X-ALMA-13B-Group{group_id}") tokenizer = AutoTokenizer.from_pretrained(f"haoranxu/X-ALMA-13B-Group{group_id}", padding_side='left') ``` **The third way**: loading the base model with all language-specific modules like MoE: (Require large GPU memory) ``` from modeling_xalma import XALMAForCausalLM model = XALMAForCausalLM.from_pretrained("haoranxu/X-ALMA", torch_dtype=torch.float16, device_map="auto") tokenizer = AutoTokenizer.from_pretrained("haoranxu/X-ALMA", padding_side='left') # Add `lang="zh"`: specify the language to instruct the model on which group to use for the third loading method during generation. generated_ids = model.generate(input_ids=input_ids, num_beams=5, max_new_tokens=20, do_sample=True, temperature=0.6, top_p=0.9, lang="zh") ``` The ALMA and ALMA-R translation prompt is: ``` Translate this from <source language name> into <target language name>: <source language name>: <source language sentence> <target language name>: ``` The X-ALMA translation prompt is: ``` <s>[INST] Translate this from <source language name> into <target language name>: <source language name>: <source language sentence> <target language name>: [/INST] ``` # Environment Setup 🔧 ``` conda create -n xalma python=3.11 conda activate xalma ``` If you use **AMD GPUs**, please first install torch with ROCm. Then install other dependencies: ``` bash install_alma.sh ``` # Evaluation 💻 ### Evaluation on X-ALMA This is a quick start to evaluate our X-ALMA model. To produce translation outputs for FLORES-200 in both en→cs and cs→en directions, (If you want to evaluate WMT'23 instead, simply pass `--override_test_data_path haoranxu/WMT23-Test`.), run the following command. **Note that You don't need enable `--chat_style` for ALMA and ALMA-R. This is only for X-ALMA** ``` accelerate launch --config_file configs/deepspeed_eval_config_bf16.yaml \ run_llmmt.py \ --model_name_or_path haoranxu/X-ALMA-13B-Group5 \ --do_predict \ --low_cpu_mem_usage \ --language_pairs en-cs,cs-en \ --mmt_data_path placeholder \ --override_test_data_path haoranxu/FLORES-200 \ --per_device_eval_batch_size 1 \ --output_dir ./your_output_dir/ \ --predict_with_generate \ --max_new_tokens 256 \ --max_source_length 256 \ --bf16 \ --seed 42 \ --num_beams 5 \ --overwrite_cache \ --overwrite_output_dir \ --chat_style # `--chat_style` only for X-ALMA. You don't need enable `--chat_style` for ALMA and ALMA-R ``` The generated outputs will be saved in the `your_output_dir`. The translation file for the `en→cs` direction is named `test-en-cs`, and the file for the cs→en direction is `test-cs-en`. The variable `${test_pairs}` denotes the translation directions you wish to evaluate. It supports testing multiple directions at once. For example, you can use `de-en,en-de,en-cs,cs-en`. Please see more other examples for evaluating ALMA(-R) under the `./evals` folder: **Note that this will perform data-parallel evaluation supported by deepspeed: that is, placing a single full copy of your model onto each available GPU and splitting batches across GPUs to evaluate on K GPUs K times faster than on one**. For those with limited GPU memory, we offer an alternative method. The user can pass `--multi_gpu_one_model` to run the process by distributing a single model across multiple GPUs. Please see evaluation examples in `evals/alma_13b_r.sh` or `evals/*no_parallel` files. # Training 🔥 Here we show how to - contrastive Preference Optmization Upon ALMA Models (ALMA→ALMA-R). - fine-tune LLaMA-2-7B on monolingual OSCAR data (stage 1) - fine-tune human-written parallel data fine-tuning once stage 1 is completed, including full-weight and LoRA fine-tuning (stage 2) Please note that we do not share the training process for X-ALMA specifically, as it would require releasing numerous intermediate checkpoints, making the process overly complex. ## **CPO Fine-Tuning** To run the CPO fine-tuning with our triplet preference data, run the following command: ``` bash runs/cpo_ft.sh ${your_output_dir} ``` ### OSCAR Monolingual Fine-Tuning To execute the OSCAR monolingual fine-tuning, use the following command: ``` bash runs/mono_ft.sh ${your_output_dir} ``` ### Parallel Data Fine-Tuning (Full-Weight) Once the monolingual data fine-tuning is complete, proceed to the parallel data fine-tuning using the full-weight approach. Execute the following command: ``` bash runs/parallel_ft.sh ${your_output_dir} $training_pairs$ ``` where `training_pairs` is the translation directions you considered. The default is all 10 directions: `de-en,cs-en,is-en,zh-en,ru-en,en-de,en-cs,en-is,en-zh,en-ru`. ### Parallel Data Fine-Tuning (LoRA) In Stage 2, there's also an option to employ LoRA for fine-tuning on the parallel data. To do so, execute the following command: ``` bash runs/parallel_ft_lora.sh ${your_output_dir} $training_pairs$ ``` # FAQs ❓ ### What language directions do ALMA and ALMA-R support? Currently, ALMA supports 10 directions: English↔German, Englishs↔Czech, Englishs↔Icelandic, Englishs↔Chinese, Englishs↔Russian. However, it may surprise us in other directions :) ### What language directions do X-ALMA support? X-ALMA supports 50 languages and 98 directions (into and from English): da,nl,de,is,no,sv,af,ca,ro,gl,it,pt,es,bg,mk,sr,uk,ru,id,ms,th,vi,mg,fr,hu,el,cs,pl,lt,lv,ka,zh,ja,ko,fi,et,gu,hi,mr,ne,ur,az,kk,ky,tr,uz,ar,he,fa ### When should I stop fine-tuning at stage 1? Our 7B and 13B models are trained on 20B and 12B tokens, respectively. However, as indicated in the paper, fine-tuning 1B tokens should boost the performance substantially. The steps required to fine-tune 1 billion tokens also vary based on your batch size. In our case, the batch size is calculated as follows: 16 GPUs * 4 (batch size per GPU) * 4 (gradient accumulation steps) = 256. With a sequence length of 512, we need approximately 8,000 steps to train on 1 billion tokens, calculated as 10^9 / (256*512) ≈8000 steps. However, you may choose to fine-tune more steps to get better performance. ### How to decide the interleave probability at stage 1? Please find the reasons for interleave probability selection for stage 1 in Appendix D.1 in the [paper](https://arxiv.org/pdf/2309.11674.pdf)! # Reference Please find more details for ALMA models in our [paper](https://arxiv.org/abs/2309.11674) or the [summary](https://notes.aimodels.fyi/alma-a-new-training-method-that-boosts-translation-performance-for-large-language-models/) of the paper. ``` @inproceedings{ xu2024a, title={A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models}, author={Haoran Xu and Young Jin Kim and Amr Sharaf and Hany Hassan Awadalla}, booktitle={The Twelfth International Conference on Learning Representations}, year={2024}, url={https://openreview.net/forum?id=farT6XXntP} } ``` Please also find more detailed information for the ALMA-R model with Contrastive Preference Optimization in the [paper](https://arxiv.org/pdf/2401.08417v2.pdf). ``` @inproceedings{ xu2024contrastive, title={Contrastive Preference Optimization: Pushing the Boundaries of {LLM} Performance in Machine Translation}, author={Haoran Xu and Amr Sharaf and Yunmo Chen and Weiting Tan and Lingfeng Shen and Benjamin Van Durme and Kenton Murray and Young Jin Kim}, booktitle={Forty-first International Conference on Machine Learning}, year={2024}, url={https://openreview.net/forum?id=51iwkioZpn} } ``` Please find details about X-ALMA in the latest [paper](https://arxiv.org/pdf/2410.03115) ``` @inproceedings{ xu2025xalma, title={X-{ALMA}: Plug \& Play Modules and Adaptive Rejection for Quality Translation at Scale}, author={Haoran Xu and Kenton Murray and Philipp Koehn and Hieu Hoang and Akiko Eriguchi and Huda Khayrallah}, booktitle={The Thirteenth International Conference on Learning Representations}, year={2025}, url={https://openreview.net/forum?id=csbf1p8xUq} } ``` ", Assign "at most 3 tags" to the expected json: {"id":"7209","tags":[]} "only from the tags list I provide: [{"id":77,"name":"3d"},{"id":89,"name":"agent"},{"id":17,"name":"ai"},{"id":54,"name":"algorithm"},{"id":24,"name":"api"},{"id":44,"name":"authentication"},{"id":3,"name":"aws"},{"id":27,"name":"backend"},{"id":60,"name":"benchmark"},{"id":72,"name":"best-practices"},{"id":39,"name":"bitcoin"},{"id":37,"name":"blockchain"},{"id":1,"name":"blog"},{"id":45,"name":"bundler"},{"id":58,"name":"cache"},{"id":21,"name":"chat"},{"id":49,"name":"cicd"},{"id":4,"name":"cli"},{"id":64,"name":"cloud-native"},{"id":48,"name":"cms"},{"id":61,"name":"compiler"},{"id":68,"name":"containerization"},{"id":92,"name":"crm"},{"id":34,"name":"data"},{"id":47,"name":"database"},{"id":8,"name":"declarative-gui "},{"id":9,"name":"deploy-tool"},{"id":53,"name":"desktop-app"},{"id":6,"name":"dev-exp-lib"},{"id":59,"name":"dev-tool"},{"id":13,"name":"ecommerce"},{"id":26,"name":"editor"},{"id":66,"name":"emulator"},{"id":62,"name":"filesystem"},{"id":80,"name":"finance"},{"id":15,"name":"firmware"},{"id":73,"name":"for-fun"},{"id":2,"name":"framework"},{"id":11,"name":"frontend"},{"id":22,"name":"game"},{"id":81,"name":"game-engine "},{"id":23,"name":"graphql"},{"id":84,"name":"gui"},{"id":91,"name":"http"},{"id":5,"name":"http-client"},{"id":51,"name":"iac"},{"id":30,"name":"ide"},{"id":78,"name":"iot"},{"id":40,"name":"json"},{"id":83,"name":"julian"},{"id":38,"name":"k8s"},{"id":31,"name":"language"},{"id":10,"name":"learning-resource"},{"id":33,"name":"lib"},{"id":41,"name":"linter"},{"id":28,"name":"lms"},{"id":16,"name":"logging"},{"id":76,"name":"low-code"},{"id":90,"name":"message-queue"},{"id":42,"name":"mobile-app"},{"id":18,"name":"monitoring"},{"id":36,"name":"networking"},{"id":7,"name":"node-version"},{"id":55,"name":"nosql"},{"id":57,"name":"observability"},{"id":46,"name":"orm"},{"id":52,"name":"os"},{"id":14,"name":"parser"},{"id":74,"name":"react"},{"id":82,"name":"real-time"},{"id":56,"name":"robot"},{"id":65,"name":"runtime"},{"id":32,"name":"sdk"},{"id":71,"name":"search"},{"id":63,"name":"secrets"},{"id":25,"name":"security"},{"id":85,"name":"server"},{"id":86,"name":"serverless"},{"id":70,"name":"storage"},{"id":75,"name":"system-design"},{"id":79,"name":"terminal"},{"id":29,"name":"testing"},{"id":12,"name":"ui"},{"id":50,"name":"ux"},{"id":88,"name":"video"},{"id":20,"name":"web-app"},{"id":35,"name":"web-server"},{"id":43,"name":"webassembly"},{"id":69,"name":"workflow"},{"id":87,"name":"yaml"}]" returns me the "expected json"