WIP: Fix Various JSON-Schema Generation Bugs

[lapp0]

Overview

The repetition problem of language models combined with patterns allowing for infinite-length fields results in broken JSON Schema outputs.

This was addressed previously for infinite whitespaces issues by setting a safe whitespace pattern as the default. In this PR, the safety of whitespaces is extended to Integer and String patterns.

Behavior

json_schema.to_regex now includes an kwarg safe_subset=True.

safe_subset=False

• Whitespace: r"[\n\t ]*"
• Integer: any number
• String: any string

safe_subset=True (default)

• Whitespace: r"[ ]?"
• Integer: (-1e19, 1e19)
• String: Any string of length (0, 256)

Fixes

Safe Integer

• Fixes llama_cpp - JSON fails to generate when using Pydantic model with models.llama_cpp dottxt-ai/outlines#1110
• Fixes <Please write a descriptive title> dottxt-ai/outlines#1099

Safe String

• Fixes The example provided by the official source has crashed. dottxt-ai/outlines#1075
• Addresses Add a non_strict mode to deal with JSONDecodeError errors dottxt-ai/outlines#985 (doesn't fix, they requested a non_strict mode)
• Fixes Accelerate not installed in Docker Image dottxt-ai/outlines#1106
  • While this technically fixes their error, their output is still bad because they didn't apply a chat template Format prompts Using Chat Templates in SequenceGeneratorAdapter dottxt-ai/outlines#987

Further Work

• Important: Low quality outputs are often caused by not applying a chat template. All examples should include chat templates, or user response quality will suffer. Format prompts Using Chat Templates in SequenceGeneratorAdapter dottxt-ai/outlines#987

• Make code more failsafe: Add a non_strict mode to deal with JSONDecodeError errors dottxt-ai/outlines#985

• number has no safe_subset implementation. Its' likely the only unsafe primitive remaining without one. There aren't any open issues for an error caused by number, however.