Update README.md

README.md

@@ -17,11 +17,14 @@ base_model:
 
 # Llama-Krikri-8B-Instruct: An Instruction-tuned Large Language Model for the Greek language
 
-Following the release of [Meltemi-7B](https://huggingface.co/ilsp/Meltemi-7B-v1) on the 26th March 2024, we are happy to welcome Krikri to the family of ILSP open Greek LLMs.
-Krikri is built on top of [Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B), extending its capabilities for Greek through continual pretraining on a large corpus of high-quality and locally relevant Greek texts. We present **Llama-Krikri-8B-Instruct**, along with the base model, [Llama-Krikri-8B-Base](https://huggingface.co/ilsp/Llama-Krikri-8B-Base).
+<div align="center">
+  <img src="https://huggingface.co/ilsp/Llama-Krikri-8B-Instruct/resolve/main/KriKri_Logo-eng_54307d80-ee25-49f9-9204-0ce774499fbc.svg?raw=true" width="60%" alt="Krikri" />
+</div>
 
+Following the release of [Meltemi-7B](https://huggingface.co/ilsp/Meltemi-7B-v1) on the 26th March 2024, we are happy to welcome Krikri to the family of ILSP open Greek LLMs.
+Krikri is built on top of [Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B), extending its capabilities for Greek through continual pretraining on a large corpus of high-quality and locally relevant Greek texts. We present **Llama-Krikri-8B-Instruct**, along with the base model, [Llama-Krikri-8B-Base](https://huggingface.co/ilsp/Llama-Krikri-8B-Base)
 
-![image/png](https://cdn-uploads.huggingface.co/production/uploads/639215a81bae0dde85842ab8/VMTgYzygHsarC9QRv2rGV.png)
+<!-- ![image/png](llama-krikri-image.jpg) -->
 
 
 # Model Information
@@ -43,7 +46,7 @@ Krikri is built on top of [Llama-3.1-8B](https://huggingface.co/meta-llama/Llama
 | English   | 21.0 B   | 23.1 %      |
 | Parallel  |  5.5 B   | 6.0 %       |
 | Math/Code |  7.8 B   | 8.6 %       |
-| **Total** | 91 B   |  **100%**       |
+| **Total** | **91 B**   |  **100%**       |
 
 
 Chosen subsets of the 91 billion corpus were upsampled resulting in a size of **110 billion tokens**.
@@ -60,7 +63,25 @@ Llama-Krikri-8B-Instruct is the result of post-training Llama-Kriki-8B-Base and
 - Conversion or structured extraction (e.g., XML, JSON) in data-to-text & text-to-data settings.
 - Analytical thinking and Chain-of-Thought (CoT) reasoning for problem-solving.
 
-🚨 **More information on the post-training corpus and methdology coming soon.** 🚨
+## Post-training Methodology
+
+We used a multi-stage process in order to build Llama-Krikri-8B-Instruct which includes:
+- 2-stage Supervised Fine-Tuning with a combination of Greek & English instruction-response pairs (& multi-turn conversations)
+  - **Stage 1**: **856,946** instruction-response pairs (371,379 Greek + 485,567 English)
+  - **Stage 2**: **638,408** instruction-response pairs (279,948 Greek + 358,460 English)
+- Alignment with a combination of Greek & English preference triplets (Instruction - Chosen Response - Rejected Response)
+  - **Length Normalized DPO**: **92,394** preference triplets (47,132 Greek + 45,262 English)
+
+## Post-training Data Construction
+
+To build the SFT & DPO data, we utilized various methodologies including:
+- Collecting existing high-quality datasets such as [Tulu 3](https://huggingface.co/datasets/allenai/tulu-3-sft-mixture), [SmolTalk](https://huggingface.co/datasets/HuggingFaceTB/smoltalk), [MAGPIE Ultra](https://huggingface.co/datasets/argilla/magpie-ultra-v1.0), [Orca Agent Instruct](https://huggingface.co/datasets/microsoft/orca-agentinstruct-1M-v1), [IFEval Like Data](https://huggingface.co/datasets/argilla/ifeval-like-data), [UltraFeedback](https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized), [NVIDIA HelpSteer2](https://huggingface.co/datasets/nvidia/HelpSteer2), [Intel Orca](https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs), [UltraMedical](https://huggingface.co/datasets/TsinghuaC3I/UltraMedical-Preference), and other datasets focused on safety, truthfulness, and instruction-following.
+- Translating various data into Greek using an in-house translation tool.
+- Regenerating translated data and contrasting the translated with the regenerated responses (i.e., for creating preference triplets).
+- Distilling (with the MAGPIE methodology) models which exhibit strong performance in Greek, such as [Gemma 2 27B IT](https://huggingface.co/google/gemma-2-27b-it).
+- Scoring data with the [Skywork Reward Gemma 2 27B v0.2](https://huggingface.co/Skywork/Skywork-Reward-Gemma-2-27B-v0.2) Reward Model and filtering using rule-based filters.
+- Creating data for sentence and document translation using high-quality parallel corpora mainly from [ELRC-SHARE](https://elrc-share.eu/).
+- Synthetically extracting question-answer pairs and multi-turn dialogues from diverse sources such as Wikipedia, EUR-LEX, Greek School Books, and Kallipos.
 
 
 # How to use
@@ -133,7 +154,43 @@ print(response.choices[0].message.content)
 
 # Evaluation
 
-🚨 **Instruction following and chat capability evaluation benchmarks coming soon.** 🚨
+In the table below, we report the scores for our chat evaluation suite which includes:
+- [Greek IFEval](https://huggingface.co/datasets/ilsp/ifeval_greek) (strict average)
+- [English IFEval](https://huggingface.co/datasets/google/IFEval) (strict average)
+- [Greek MT-Bench](https://huggingface.co/datasets/ilsp/mt-bench-greek) using gpt-4o-2024-08-06 as the judge model.
+- [English MT-Bench](https://huggingface.co/datasets/HuggingFaceH4/mt_bench_prompts) using gpt-4o-2024-08-06 as the judge model.
+
+We can observe that *Llama-Krikri-8B-Instruct exhibits the strongest performance* in instruction following for both Greek and English across all the models we tested. In particular, it surpasses Llama-3.1-8B-Instruct by **+21.7%** and **+7.3%** on the Greek and English IFEval respectively.
+It also exhibits **the strongest chat capabilities in the Greek MT-Bench benchmark** (+0.28 compared to Aya Expanse 8B), while also being very competitive in the English variant of the MT-Bench benchmark.
+
+|       | IFEval EL (strict avg) | IFEval EN (strict avg) | MT-Bench EL | MT-Bench EN |
+|---------------- |---------------- |----------------- |------------|------------|
+| Qwen 2.5 7B Instruct | 46.2% | 74.8% | 5.83 | **7.87** |
+| EuroLLM 9B Instruct | 51.3% | 64.5% | 5.98 | 6.27 |
+| Aya Expanse 8B | 50.4% | 62.2% | 7.68 | 6.92 |
+| Meltemi 7B v1.5 Instruct | 32.7% | 41.2% | 6.25 | 5.46 |
+| Llama-3.1-8B Instruct | 45.8% | 75.1% | 6.46 | 7.25 |
+| **Llama-Krikri-8B Instruct** | **67.5%** | **82.4%** | **7.96** | 7.21 |
+
+
+We also used the [Arena-Hard-Auto](https://huggingface.co/datasets/lmarena-ai/arena-hard-auto-v0.1) automatic evaluation tool, as well the translated (and post-edited) version for Greek that is publicly available [here](https://huggingface.co/datasets/ilsp/m-ArenaHard_greek). We report 2 scores for Arena-Hard-Auto:
+- No Style Control: The original version of the benchmark.
+- With Style Control: The benchmark with style control methods for Markdown elements. You can read more about the methodology and technical background in this [blogspot](https://lmsys.org/blog/2024-08-28-style-control/).
+
+Below, we show the scores for the Greek version of Arena-Hard-Auto for various open and closed chat models that were determined using **gpt-4o-2024-08-06 as the judge model** and **gpt-4o-mini-2024-07-18 as the baseline model** (i.e., by default 50% score).
+
+Llama-Krikri-8B Instruct exhibits very strong chat capabilities by scoring **higher than models over 8 times its size** (such as Llama-3.1-70B Instruct) and is also **competitive with closed-source** (e.g., GPT-4o-Mini) and **highly-performant open-source models** (e.g., Gemma 2 27B IT & Aya Expanse 32B). 
+![image/png](arena_hard_el.png)
+
+Below, we show the scores for the original Arena-Hard-Auto dataset for various open and closed chat models. We followed the original methodology by using **gpt-4-1106-preview as the judge model** and **gpt-4-0314 as the baseline model**.
+
+Llama-Krikri-8B Instruct performs very well in the English variant of Arena-Hard-Auto as well, since we can observe that it is **competitive with similarly sized LLMs** and that it **improves upon Llama-3.1-8B Instruct by +24.5% / +16%** (No style control / With style control). 
+![image/png](arena_hard_en.png)
+
+***Please note** that judge models are biased towards student models trained on distilled data from them. You can read more [here](https://arxiv.org/pdf/2502.01534?).
+
+🚨 **More information on post-training, methodology, and evaluation coming soon.** 🚨
+
 
 # Acknowledgements