ArcLM

On ArcLM 2.0.0, using the given tiny model configuration, the output logits shape on CPU is (1, 3, 11), and the parameter count is 3531.

The ArcLM core installation automatically pulls in dependencies such as NumPy and SentencePiece, but does not pull PyTorch.

In the tutorial, the ArcLM version output is 2.0.0.

ArcLM has a built-in simple word-level tokenizer.

On ArcLM 2.0.0, using max_vocab=64 and the given text, the tokenization output Token IDs are [2, 3, 4], the vocabulary size is 11, and it decodes to ['arc', 'lm', 'trains'].

The ArcLM tokenizer does not automatically lowercase, 'Arc' and 'arc' are treated as different Tokens, and words outside the vocabulary are mapped to an unknown Token.

ArcLM provides both a word-level tokenizer and SentencePiece support.

In ArcLM 2.0, configuration and execution are separated, allowing model settings to be defined without installing PyTorch.

ArcLM's configuration can be saved as YAML and loaded.

ArcLM is a Python framework focused on language model toolchains.

ArcLM version 2.0 separates lightweight tools from heavy computation functionality, and the core package does not mandatorily depend on PyTorch.

ArcLM's ToolPermission.PURE indicates that a tool does not need to access external resources, and SideEffectLevel.NONE indicates that a tool has no side effects.

SideEffectLevel describes whether the impact a tool has on the external world after execution can be undone, with common levels including NONE (pure computation), REVERSIBLE (undoable, such as writing to a rollback-capable temporary file), and irreversible or externally visible operations.

In ArcLM, permission and side_effect are two mutually independent dimensions: permission answers what resources a tool needs to access, and side effect answers what consequences are left after access.

ArcLM's tool abstraction performs parameter validation before calling the underlying function.

ArcLM's model adapter is responsible for converting the framework's internal sessions and tool definitions into the request format required by the LLM vendor API, and converting the returned results into a ModelResponse structure; it typically needs to implement a method like generate_response(messages, tools, instructions, session).

The DeterministicFakeModel in this article's demonstration is a fake model adapter that ignores model inference and returns preset tool calls or final replies according to the call count, for testing the complete tool-calling pipeline under conditions with no network and no GPU.

ArcLM's design allows that when replacing a fake model with a real model, only the adapter needs to be replaced, with no need to rewrite tools and business code.

ArcLM is a Python framework/library for building local Agent tool calls, providing layered components such as tool definitions, a registry, adapters, a runner, and permission redaction.