Update README.md

README.md

@@ -1,15 +1,172 @@
 ---
-license: llama2
 language:
 - nl
+license: llama2
 ---
 
-## LLaMA-2-NL: Fine-tuned using LoRa and a translated tokenizer
+<p align="center" style="margin:0;padding:0">
+<img src="./chocollama_logo.png" alt="ChocoLlama logo" width="500" style="margin-left:'auto' margin-right:'auto' display:'block'"/>
+</p>
+<div style="margin:auto; text-align:center">
+<h1 style="margin-bottom: 0">ChocoLlama</h1>
+<em>A Llama-2/3-based family of Dutch language models</em>
+</div>
 
-```
+## ChocoLlama-2-7B-tokentrans-base: Getting Started
+
+We here present **ChocoLlama-2-7B-tokentrans-base**, a language-adapted version of Meta's Llama-2-7b. 
+We replace the original tokenizer by a RoBERTa-based tokenizer trained only on Dutch data. 
+The token embeddings of the original Llama-2 model were then reinitialized using the token translation algorithm proposed by [Remy et al.](https://arxiv.org/pdf/2310.03477). 
+The model was subsequently fine-tuned on 32B Dutch Llama-2 tokens (104GB) using LoRa. 
+Using the new, Dutch-specific tokenizer, the total number of tokens is reduced to 22.6B, or a reduction of 29.4% compared to the original Llama-2 tokenizer.
+Note that this is a base model, not optimized for conversational behavior. 
+If this is desired for your use-case, we recommend finetuning this model on your own Dutch data or using the instruction-finetuned version of this model, [ChocoLlama-2-7B-tokentrans-instruct](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-tokentrans-instruct). 
+
+Use the code below to get started with the model.
+
+```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-# take the new tokenizer
-tokenizer = AutoTokenizer.from_pretrained('llama-2-nl/Llama-2-7b-hf-lora-tokentrans')
-model = AutoModelForCausalLM.from_pretrained('llama-2-nl/Llama-2-7b-hf-lora-tokentrans')
-```
+tokenizer = AutoTokenizer.from_pretrained('ChocoLlama/ChocoLlama-2-7B-tokentrans-base')
+model = AutoModelForCausalLM.from_pretrained('ChocoLlama/ChocoLlama-2-7B-tokentrans-base')
+```
+
+## Model Details
+
+ChocoLlama is a family of open LLM's specifically adapted to Dutch, contributing to the state-of-the-art of Dutch open LLM's in their weight class.
+
+We provide 6 variants (of which 3 base and 3 instruction-tuned models):
+- **ChocoLlama-2-7B-base** ([link](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-base)): A language-adapted version of Meta's Llama-2-7b, fine-tuned on 32B Dutch Llama-2 tokens (104GB) using LoRa.
+- **ChocoLlama-2-7B-instruct** ([link](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-instruct)): An instruction-tuned version of ChocoLlama-2-7B-base, fine-tuned on a collection of Dutch translations of instruction-tuning datasets, using SFT followed by DPO.
+- **ChocoLlama-2-7B-tokentrans-base** ([link](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-tokentrans-base)): A language-adapted version of Meta's Llama-2-7b, using a Dutch RoBERTa-based tokenizer. The token embeddings of this model were reinitialized using the token translation algorithm proposed by [Remy et al.](https://arxiv.org/pdf/2310.03477). The model was subsequently fine-tuned on the same Dutch dataset as ChocoLlama-2-7B-base, again using LoRa.
+- **ChocoLlama-2-7B-tokentrans-instruct** ([link](https://huggingface.co/ChocoLlama/ChocoLlama-2-7B-tokentrans-instruct)): An instruction-tuned version of ChocoLlama-2-7B-tokentrans-base, fine-tuned on the same dataset as ChocoLlama-2-7B-instruct, again using SFT followed by DPO.
+- **Llama-3-ChocoLlama-8B-base** ([link](https://huggingface.co/ChocoLlama/Llama-3-ChocoLlama-8B-base)): A language-adapted version of Meta's Llama-8-8B, fine-tuned on the same Dutch dataset as ChocoLlama-2-7B-base, again using LoRa.
+- **Llama-3-ChocoLlama-instruct** ([link](https://huggingface.co/ChocoLlama/Llama-3-ChocoLlama-8B-instruct)): An instruction-tuned version of Llama-3-ChocoLlama-8B-base, fine-tuned on the same dataset as ChocoLlama-2-7B-instruct, again using SFT followed by DPO.
+
+For benchmark results for all models, including compared to their base models and other Dutch LLMs, we refer to our paper [here](some_url).  
+
+### Model Description
+
+- **Developed by:** [Matthieu Meeus](https://huggingface.co/matthieumeeus97), [Anthony Rathé](https://huggingface.co/anthonyrathe)
+- **Funded by:** [Vlaams Supercomputer Centrum](https://www.vscentrum.be/), through a grant of apx. 40K GPU hours (NVIDIA A100-80GB)
+- **Language(s):** Dutch
+- **License:** [Llama-2 Community License](https://ai.meta.com/llama/license/)
+- **Finetuned from model:** [Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf)
+
+### Model Sources
+
+- **Repository:** Will be released soon.
+- **Paper:** Will be released soon.
+
+## Uses
+
+### Direct Use
+
+Since this is a base model, we do not recommend using it for your use-cases directly. We instead recommend:
+1. Fine-tuning this model to your specific use-case
+2. Leveraging the instruction-tuned version of this model
+
+### Downstream Use
+
+Since this model is a base model, it can easily be adapted to specific use-cases that required Dutch language understanding and generation. 
+We expect this model to be particularly useful for use-cases in the domains which were explicitly covered in our dataset, e.g. the analysis and/or generation of Dutch job descriptions, corporate filings and legislation.
+
+### Out-of-Scope Use
+
+- Use-cases requiring a chat-style interface: since this is a base model, it cannot be used reliably for turn-based chat interaction. Please refer to the instruction-tuned version of this model instead.
+- Use-cases requiring understanding or generation of text in languages other than Dutch: the dataset on which this model was fine-tuned does not contain data in languages other than Dutch, hence we expect significant catastrophic forgetting to have occured for English, which is the language Llama-2 was originally trained for.
+
+## Bias, Risks, and Limitations
+
+We have taken care to include only widely used and high-quality data in our dataset. Some of this data has been filtered by the original creators.
+However we did not explicitly conduct any additional filtering of this dataset with regards to biased or otherwise harmful content.
+
+### Recommendations
+
+We recommend fine-tuning this model to your curated data to maximally avoid undesirable outputs.
+
+## Training Details
+
+### Training Data
+
+We collect a diverse set of Dutch natural language. 
+
+1. **OSCAR**  
+   The bulk of our data comes from the Dutch portion of [OSCAR](https://oscar-corpus.com), January 2023 version, based on Common Crawl. This dataset includes **93 GB** of text (~28.6B tokens).
+
+2. **Open Subtitles**  
+   We collected Dutch text from movie subtitles, focusing on unique movies either in Dutch or with Dutch subtitles. This dataset contains **5 GB** of text (~1.54B tokens) from **214k samples**.
+
+3. **Project Gutenberg**  
+   We downloaded **970 full Dutch books** from [Project Gutenberg](https://www.gutenberg.org) using a public scraper. The dataset includes **0.3 GB** of text (~92M tokens) and is available on [Hugging Face](https://huggingface.co/datasets/ChocoLlama/gutenberg-dutch).
+
+4. **Wikipedia**  
+   Using the March 2023 [Wikipedia dump](https://dumps.wikimedia.org), we included **2.5 GB** of text (~769M tokens). Despite some duplication with OSCAR, Wikipedia's high quality justifies its inclusion.
+
+5. **Job Descriptions (TechWolf)**  
+   A sample of **750k Dutch job descriptions** collected over five years from public websites, provided by TechWolf. This dataset contains **1.5 GB** of text (~462M tokens).
+
+6. **Staatsblad (Bizzy)**  
+   A sample of **80k legal filings** from [Het Belgisch Staatsblad](https://www.ejustice.just.fgov.be/cgi/welcome.pl). Documents were OCR-processed, and personal data was excluded. This dataset includes **1.4 GB** of text (~431M tokens), collected with help from Bizzy.
+
+7. **Legislation (ML6)**  
+   **15k documents** from Flemish legislation accessed via the [Open Data API](https://www.vlaanderen.be/vlaams-parlement/de-vlaamse-codex). This dataset contains **0.2 GB** of text (~62M tokens), collected with support from ML6.
+
+### Training Procedure
+
+This model was fine-tuned using low-rank (LoRa) adapatation with trainable embeddings, for a total of 839M trainable parameters.
+
+#### Training Hyperparameters
+
+- **Training regime:** bf16 non-mixed precision
+- **Epochs:** 1
+- **LoRa parameters:**
+    - R: 8
+    - Alpha: 32
+    - Trainable modules: q_proj, v_proj, k_proj, o_proj, gate_proj, up_proj, down_proj, embed_tokens, lm_head
+    - LoRa dropout: 0.05
+- **Learning Rate:**
+    - Scheduler: StepLR
+    - Step size: 6212
+    - Learning rate: 0.0003
+    - Gamma: 0.85
+- **Other parameters:**
+    - Minibatch size: 16
+    - Gradient accumulation steps: 8
+    - Parallelization factor: 8
+    - Weight decay: 0
+
+## Evaluation
+
+### Quantitative evaluation
+
+We have evaluated our models on several industry-standard Dutch benchmarks, translated from their original versions. The results can be found in the table below, together with results from several other prominent Dutch models.
+
+| Model                                        | ARC            | HellaSwag      | MMLU           | TruthfulQA     | Avg.           |
+|----------------------------------------------|----------------|----------------|----------------|----------------|----------------|
+| **Llama-3-ChocoLlama-instruct**        | **0.48** | **0.66** | **0.49** | **0.49** | **0.53** |
+| llama-3-8B-rebatch                           | 0.44           | 0.64           | 0.46           | 0.48           | 0.51           |
+| llama-3-8B-instruct                          | 0.47           | 0.59           | 0.47           | 0.52           | 0.51           |
+| llama-3-8B                                   | 0.44           | 0.64           | 0.47           | 0.45           | 0.5            |
+| Reynaerde-7B-Chat                            | 0.44           | 0.62           | 0.39           | 0.52           | 0.49           |
+| **Llama-3-ChocoLlama-base** | **0.45** | **0.64** | **0.44** | **0.44** | **0.49** |
+| zephyr-7b-beta                               | 0.43           | 0.58           | 0.43           | 0.53           | 0.49           |
+| geitje-7b-ultra                              | 0.40           | 0.66           | 0.36           | 0.49           | 0.48           |
+| **ChocoLlama-2-7B-tokentrans-instruct** | **0.45** | **0.62** | **0.34** | **0.42** | **0.46** |
+| mistral-7b-v0.1                              | 0.43           | 0.58           | 0.37           | 0.45           | 0.46           |
+| **ChocoLlama-2-7B-tokentrans-base** | **0.42** | **0.61** | **0.32** | **0.43** | **0.45** |
+| **ChocoLlama-2-7B-instruct** | **0.36** | **0.57** | **0.33** | **0.45** | **0.43 |
+| **ChocoLlama-2-7B-base** | **0.35** | **0.56** | **0.31** | **0.43** | **0.41** |
+| llama-2-7b-chat-hf                           | 0.36           | 0.49           | 0.33           | 0.44           | 0.41           |
+| llama-2-7b-hf                                | 0.36           | 0.51           | 0.32           | 0.41           | 0.40           |
+
+On average, Llama-3-ChocoLlama-instruct surpasses the previous state-of-the-art on these benchmarks.
+
+### Qualitative evaluation
+
+In our paper, we also provide an additional qualitative evaluation of all models - which we empirically find more reliable. 
+For details, we refer to the paper and to our benchmark [ChocoLlama-Bench](https://huggingface.co/datasets/ChocoLlama/ChocoLlama-Bench). 
+
+### Compute Infrastructure
+
+All ChocoLlama models have been trained on the compute cluster provided by the [Flemish Supercomputer Center (VSC)](https://www.vscentrum.be/). We used 8 to 16 NVIDIA A100 GPU's with 80 GB of VRAM.