zils-lite-dh18
a weight merge for the Zils subnet
1 min readupdated
This checkpoint is a merged model (a model soup) built for Bittensor subnet 127. Based on the Qwen/Qwen3.5-0.8B vision-language model (0.8 billion parameters), it combines three fine-tuned modules using task arithmetic (summing weight deltas). It is designed for specialized text and analytical tasks within the Zils ecosystem.
Key features
- Qwen3.5 base: Built on the compact Qwen/Qwen3.5-0.8B model inside the d3-lite chassis.
- Task arithmetic: Merges a language LoRA (weight 1.0), a tool-use LoRA (0.7), and a reasoning LoRA (0.5).
- Soup-trained adapters: Each LoRA (rank 64, lr 5e-5) was trained on the initial soup using a response-mixing objective.
- Apache-2.0 license: Distributed under the terms of the original parent models and datasets.
What you can do
- Pipeline integration: Use the model as a component for text processing and tool-calling workflows.
- Bittensor deployment: Run the merge inside subnet 127 infrastructure to handle requests formatted for the Jev Decision Index.
Example
Training or inference for these components is launched through the dedicated zils CLI scripts:python -m zils.teacher train
Read more
DetailsArchitecture: Qwen/Qwen3.5-0.8B (0.8B parameters)
Merge method: task arithmetic on the daily-network d3-lite soup
Components: language, tools, and reasoning
License: Apache-2.0










Comments