OpenLLM-Ro/RoMistral-7b-Instruct-DPO-2025-04-23

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedPublished:Mar 16, 2025License:cc-by-nc-4.0Architecture:Transformer Open Weights Featherless Exclusive Cold

OpenLLM-Ro/RoMistral-7b-Instruct-DPO-2025-04-23 is a 7 billion parameter instruction-tuned generative text model developed by OpenLLM-Ro, specifically optimized for the Romanian language. This model is a human-aligned instruct variant of the RoMistral family, fine-tuned using various Romanian DPO datasets. It excels in Romanian natural language tasks, demonstrating strong performance across academic benchmarks like ARC, MMLU, and Hellaswag, and achieving top scores in Romanian-specific evaluations such as MT-Bench and RoCulturaBench.

Loading preview...

OpenLLM-Ro/RoMistral-7b-Instruct-DPO-2025-04-23 Overview

This model is a 7 billion parameter, instruction-tuned generative text model developed by OpenLLM-Ro, specifically designed for the Romanian language. It represents a significant open-source effort to create a specialized LLM for Romanian, building upon the RoMistral foundational models. The model is a human-aligned instruct variant, fine-tuned using a collection of Romanian Direct Preference Optimization (DPO) datasets including RoHelpSteer, RoUltraFeedback, RoMagpieDPO, RoArgillaMagpie, and RoHelpSteer2.

Key Capabilities and Performance

  • Romanian Language Specialization: Optimized for natural language tasks exclusively in Romanian, making it highly effective for local applications.
  • Instruction Following: Designed to function as an assistant-like chat model, capable of understanding and responding to instructions.
  • Strong Benchmark Results: Achieves an average score of 56.62 on academic benchmarks, outperforming other RoMistral variants and Mistral-7B-Instruct-v0.2 in several categories, including ARC (55.51), MMLU (52.61), Hellaswag (64.97), and GSM8k (41.07).
  • High Romanian Fluency: Demonstrates superior performance in Romanian-specific evaluations, scoring 6.61 on MT-Bench (with 100% answers in Romanian) and 4.93 on RoCulturaBench (with 100% answers in Romanian).
  • Downstream Task Proficiency: Shows competitive results in few-shot and finetuned scenarios for tasks like LaRoSeDa (Binary Macro F1: 97.94), WMT (EN-RO Bleu: 27.24), and XQuAD (EM: 40.86, F1: 62.24).

Intended Use Cases

  • Research in Romanian NLP: Ideal for academic and research purposes focused on the Romanian language.
  • Assistant-like Chatbots: Suitable for developing conversational AI agents that interact in Romanian.
  • Natural Language Tasks: Can be adapted for various Romanian natural language processing tasks where instruction following is crucial.

Finetunes
1 models