LLST
An open-source standardized evaluation tool for local large language models, comprising two layers: a fixed 102-question capability benchmark covering knowledge reasoning, instruction following, advanced mathematics, Chinese language ability, and code generation; and a deterministic long-context stress test supporting cross-model comparability.
Timeline (last 90 days)
LLST (Local Large Model Standard Test) is an open-source standardized local large model testing tool consisting of two components: a capability benchmark and a gradient stress test.
LLST's capability test uses a fixed set of 102 prompt samples, covering five dimensions: knowledge reasoning, instruction following, advanced mathematics, Chinese language ability, and code generation.
The distribution of questions across LLST's five dimensions is as follows: 42 for knowledge reasoning, 20 for instruction following, 10 for advanced mathematics, 20 for Chinese language ability, and 10 for code generation.
LLST's stress test fixes four input gradient levels of 512, 4K, 16K, and 28K tokens, with output uniformly fixed at 512 tokens.
LLST interfaces via the token mode of the OpenAI-compatible API, supporting multiple inference frameworks.
The Mac client software for LLST is not independently developed, but is a secondary development based on the publicly available free tool UnaScope.
All Facts (6)
LLST (Local Large Model Standard Test) is an open-source standardized local large model testing tool consisting of two components: a capability benchmark and a gradient stress test.
50%UnverifiedLLST's capability test uses a fixed set of 102 prompt samples, covering five dimensions: knowledge reasoning, instruction following, advanced mathematics, Chinese language ability, and code generation.
50%UnverifiedThe distribution of questions across LLST's five dimensions is as follows: 42 for knowledge reasoning, 20 for instruction following, 10 for advanced mathematics, 20 for Chinese language ability, and 10 for code generation.
50%UnverifiedLLST's stress test fixes four input gradient levels of 512, 4K, 16K, and 28K tokens, with output uniformly fixed at 512 tokens.
50%UnverifiedLLST interfaces via the token mode of the OpenAI-compatible API, supporting multiple inference frameworks.
50%UnverifiedThe Mac client software for LLST is not independently developed, but is a secondary development based on the publicly available free tool UnaScope.
50%