README.md

metadata

license: apache-2.0

Run fix-json on Synthetic! IMPORTANT: Make sure to set temperature to zero, either via the API or in the settings UI on Synthetic (accessed via the gear icon).

fix-json is a small, fast model to fix broken JSON. While most LLM inference providers support a "JSON mode" that enforces model responses are valid JSON, many don't support strict JSON or schema adherence for function calls. While coding LLMs are typically fairly good at generating JSON, at a certain level of nesting they can make small mistakes like forgetting to triple-escape some deeply nested control character.

When that happens, most agent frameworks force retries. Retries are bad for a couple of reasons:

1. They're slow, and typically expensive, since you're using a fairly large model to do the retry
2. They often can confuse the model: models are trained reasonably well on the happy path where they generate working code and tool calls, but often suffer when attempting to self-correct.

fix-json is trained to fix those broken JSON strings, allowing coding agents to typically skip (expensive, slow, model-confusing) retries, fairly robustly. It's a Llama 3.1 8b LoRA, which means you can run it cheaply on inference providers like Synthetic. It was trained against editing any inner string contents, and attempts to only make the minimal control character edits possible to fix the JSON string.

The model is generally fairly robust to prompts, but for best performance you should use the original prompt it was trained on:

The following string may be broken JSON. Fix it if possible. Respond with JSON in the following
format, defined as TypeScript types:

// Success response:
type JsonFixSuccess = {
  success: true,

  // The parsed JSON
  fixed: any,
};

// Failure response:
type JsonFixFailure = {
  success: false,
};

If it's more-or-less JSON, fix it and respond with the success response. If it's not, respond with
the failure response. Here's the string:

You should append the broken JSON string to the end of the prompt. It'll respond with either { "success": false } if the string is too ambiguous or doesn't appear to be JSON; otherwise, it'll generate { "success": true, "fixed": ... }, where fixed is the JSON.

Since the model responds with pure JSON, it's suitable for use with LLM inference providers that allow JSON mode, making it even more robust at fixing JSON.

Training details

Trained on a single H100 NVL for 2.5 hours using Axolotl.