bond005/meno-tiny-0.1

Warm
Public
1.5B
BF16
32768
Nov 18, 2024
License: apache-2.0
Hugging Face
Overview

Meno-Tiny-0.1: A Russian-Optimized Language Model

Meno-Tiny-0.1 is a 1.5 billion parameter language model, a fine-tuned descendant of Qwen2.5-1.5B-Instruct, developed by Ivan Bondarenko. Its name, "Meno," reflects its specialization in question answering from text, particularly within Retrieval-Augmented Generation (RAG) pipelines. The model leverages a Transformer architecture with features like SwiGLU activation and group query attention.

Key Capabilities

  • Russian Language Proficiency: Specifically "Russified" through fine-tuning on a dedicated Russian instruct dataset, while retaining English language capabilities.
  • Diverse NLP Tasks: Excels at a range of tasks including:
    • Answering questions about text
    • Text summarization
    • Determining and detoxifying text toxicity
    • Anaphora resolution in dialogues
    • Correcting speech recognition errors
  • Performance on Russian Benchmarks: Achieves a 0.365 overall score on the MERA benchmark, an independent evaluation for Russian language models, outperforming its base model. It also scores 0.399 / 0.29 on the MultiQ task, crucial for RAG effectiveness.

Good For

  • Memory/Compute-Constrained Environments: Designed for efficient operation in resource-limited settings.
  • Latency-Bound Scenarios: Suitable for applications requiring quick response times.
  • Retrieval Augmented Generation (RAG): Serves as a strong building block for RAG pipelines, particularly for Russian-language content.
  • Research and Commercial Use: Intended for both research and commercial applications requiring robust Russian NLP capabilities.