Kwaipilot/KAT-Dev-72B-Exp
Kwaipilot/KAT-Dev-72B-Exp is an experimental 72-billion parameter model developed by Kwaipilot, specifically designed for software engineering tasks. This model is the reinforcement learning version of the KAT-Coder, achieving 74.6% accuracy on SWE-Bench Verified when evaluated with the SWE-agent scaffold. It incorporates technical innovations in attention kernel rewriting and a redesigned training engine for efficient RL, particularly for context-managed scaffolds. Its primary strength lies in advanced code generation and problem-solving within software development environments.
Hugging Face Popular Sampler Settings
Most commonly used values from Featherless users
temperature
This setting influences the sampling randomness. Lower values make the model more deterministic; higher values introduce randomness. Zero is greedy sampling.
–
top_p
This setting controls the cumulative probability of considered top tokens. Must be in (0, 1]. Set to 1 to consider all tokens.
–
top_k
This limits the number of top tokens to consider. Set to -1 to consider all tokens.
–
frequency_penalty
This setting penalizes new tokens based on their frequency in the generated text. Values > 0 encourage new tokens; < 0 encourages repetition.
–
presence_penalty
This setting penalizes new tokens based on their presence in the generated text so far. Values > 0 encourage new tokens; < 0 encourages repetition.
–
repetition_penalty
This setting penalizes new tokens based on their appearance in the prompt and generated text. Values > 1 encourage new tokens; < 1 encourages repetition.
–
min_p
This setting representing the minimum probability for a token to be considered relative to the most likely token. Must be in [0, 1]. Set to 0 to disable.
–