BangorAI/Mistral-7B-Cymraeg-Welsh-v2

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7BQuant:FP8Context Size:4kTool Calling:SupportedPublished:Mar 26, 2024License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

BangorAI/Mistral-7B-Cymraeg-Welsh-v2 is a 7 billion parameter Mistral-based chat/instruct model developed by BangorAI, continually pre-trained and fine-tuned for bilingual capabilities in both English and Welsh. It leverages the Mistral-7B-v0.1 architecture and was fine-tuned on the alpaca-cleaned dataset in both languages. This model is specifically designed for conversational tasks requiring fluency in both English and Welsh, making it ideal for applications targeting Welsh-speaking communities.

Loading preview...

Overview

BangorAI/Mistral-7B-Cymraeg-Welsh-v2 is a 7 billion parameter bilingual chat/instruct model built upon the Mistral-7B-v0.1 architecture. Developed by BangorAI, this model has undergone significant training to achieve proficiency in both English and Welsh.

Key Capabilities

  • Bilingual Fluency: Trained and fine-tuned in both English and Welsh, enabling seamless conversation in either language.
  • Instruction Following: Functions as an instruct model, capable of responding to user prompts and system instructions.
  • Llama-2 Prompt Format: Utilizes the Llama-2 prompt format, where the system prompt's language guides the model's response language.

Training Details

The model's development involved two main stages:

  1. Continual Pre-training: Based on BangorAI/mistral-7b-cy-epoch-2, which itself was a continual pre-training of Mistral-7B-v0.1 using Welsh data from the allenai/MADLAD-400 dataset for two epochs.
  2. Fine-tuning: Further fine-tuned using the yahma/alpaca-cleaned dataset, incorporating both Welsh and English data for an additional two epochs.

Use Cases

This model is particularly well-suited for applications requiring a conversational AI that can interact effectively in both English and Welsh. It's an experimental LLM, and users should exercise caution regarding factual accuracy.

An online demo is available at https://demo.bangor.ai.