metadata

license: apache-2.0
base_model:
  - NovaSky-AI/Sky-T1-32B-Preview
library_name: peft
datasets:
  - shakedzy/Sky-T1_data_steps

Sky-T1-32B with Step-by-Step Thinking

This repo contains a LoRA adapter to NovaSky-AI/Sky-T1-32B-Preview for step-by-step thinking. It breaks down the thinking process to steps, allowing better tracking of the thought process.

The model was finetuned without any system prompt.

Each step is outputted in a specific format:

<|step|> TITLE
TEXT

Thus the model's response should adhere to the following format:

<|begin_of_thought|>
<|step|> STEP_1_TITLE
STEP_1_THINKING
<|step|> STEP_2_TITLE
STEP_2_THINKING
...
<|end_of_thought|>
<|begin_of_solution|>
...
<|end_of_solution|>