Sao10K/MN-12B-Lyra-v2a1

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:12BQuant:FP8Ctx Length:32kPublished:Aug 12, 2024License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Warm

Sao10K/MN-12B-Lyra-v2a1 is a 12 billion parameter language model, representing Stage 1 of the Lyra v2 development. This model has undergone Supervised Fine-Tuning (SFT) and a 2-step merge process. It is designed for general language generation tasks, though it may exhibit verbosity similar to Claude models, which is planned for refinement in future stages. Users should note potential `[/INST]` token leakage and utilize ChatML prompt formatting.

Loading preview...

Overview

Sao10K/MN-12B-Lyra-v2a1 is the initial release (Stage 1) of the Lyra v2 model series, developed by Sao10K. This version has been processed through Supervised Fine-Tuning (SFT) and a 2-step merging procedure. It is presented as a raw model, indicating ongoing development and potential for refinement in subsequent stages.

Key Characteristics

  • Development Stage: Stage 1 of Lyra v2, with further development (including RL) planned.
  • Current Behavior: May exhibit verbose and wordy output, similar to Claude models, which is an identified area for improvement in Stage 2.
  • Prompt Format: Utilizes the ChatML prompt format for interaction.
  • Known Issues: The [/INST] token may occasionally leak into responses; users are advised to add it to custom stopping strings.

Usage Recommendations

  • min_p Value: A min_p value of 0.1 is suggested for stable generation.
  • Temperature: A temperature range of 0.7-1.2 is recommended, allowing for user preference in output creativity.

Popular Sampler Settings

Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.

temperature
top_p
top_k
frequency_penalty
presence_penalty
repetition_penalty
min_p