vizgg32dx/abyss-tiny-4b

Abyss Tiny (4B) v2 Overview

Abyss Tiny (4B) v2 is a 4 billion parameter language model developed by Kainos Labs as part of the Abyss project. It is built upon Qwen/Qwen3-4B-Base and has been fine-tuned using LoRA to enhance its capabilities in both Spanish and English.

Key Capabilities & Features

• Multilingual Proficiency: Reasons, programs, solves mathematical problems, and converses effectively in both Spanish and English.
• Tool-Use: Capable of utilizing external tools for enhanced functionality.
• Unconditional Identity: Unlike its v1 predecessor, Abyss Tiny v2 identifies itself as "Abyss Tiny" even without a specific system prompt, with this identity embedded directly into its weights.
• Improved Coherence: With 4B parameters (up from 1.7B in v1), it offers better conversational flow, reasoning, and code generation.
• ChatML Format: Supports ChatML for chat interactions, integrated into its tokenizer and GGUF versions.

Training Details

The model was fine-tuned using LoRA (r=64) on approximately 289,000 samples. The training data included a diverse mix of reasoning tasks, conversational data (SmolTalk), secure code (CodeAlpaca), mathematics with Chain-of-Thought, and tool-use examples. A significant portion of the training (~40%) focused on embedding the "Abyss Tiny" identity without requiring a system prompt.

Use Cases

This model is suitable for applications requiring strong performance in:

• General-purpose conversational AI
• Code generation and understanding
• Mathematical problem-solving
• Reasoning tasks
• Tool-augmented language model applications

It is provided as an educational and experimental model by Kainos Labs.