Model Overview

Magnum-v3-34b is a 34 billion parameter language model developed by anthracite-org, building upon the Yi-1.5-34B-32K base. This iteration, the ninth in its series, is specifically fine-tuned to emulate the prose quality of Claude 3 models, including Sonnet and Opus, focusing on nuanced and high-quality text generation.

Key Capabilities & Features

• Claude 3 Prose Replication: Optimized to produce text with a similar style and quality to Claude 3 Sonnet and Opus models.
• Instruction-Tuned: Utilizes ChatML formatting for clear instruction following, with specific templates provided for platforms like SillyTavern.
• Context Length: Supports a substantial context window of 16384 tokens, enabling more extensive and coherent interactions.
• Training: Fine-tuned over two epochs using 8x H100 GPUs, leveraging a diverse dataset including anthracite-org/stheno-filtered-v1.1 and anthracite-org/kalo-opus-instruct-22k-no-refusal.

Performance Highlights

Evaluations on the Open LLM Leaderboard show an average score of 29.39. Notable scores include:

• IFEval (0-Shot): 51.15
• BBH (3-Shot): 44.33
• MMLU-PRO (5-shot): 41.69

Recommended Use Cases

• Creative Writing: Ideal for generating detailed narratives, character dialogues, and descriptive prose.
• Roleplay Scenarios: Excels in interactive storytelling and maintaining consistent character personas.
• Conversational AI: Suitable for applications requiring human-like, high-quality conversational responses.

For optimal performance, a min_p of 0.2 is recommended during inference.