mudler/LocalAI · error · ValueError

Inline reward function '{name}' rejected: inline reward code

Error message

Inline reward function '{name}' rejected: inline reward code executes arbitrary Python and is disabled by default. Set {ALLOW_INLINE_ENV}=true on the backend to enable it (only on a trusted, access-controlled instance), or use a builtin reward function instead.

What it means

Inline reward specs execute arbitrary caller-supplied Python via exec, and the backend deliberately treats its _SAFE_BUILTINS allowlist as non-security (comments note trivial escapes via __subclasses__). Because the fine-tuning endpoint is unauthenticated by default, inline rewards are opt-in: the backend refuses them unless LOCALAI_TRL_ALLOW_INLINE_REWARD is set to a truthy value (1/true/yes/on).

Source

Thrown at backend/python/trl/reward_functions.py:242

                )
            func = BUILTIN_REGISTRY[name]
            if params:
                # Convert string params to appropriate types
                typed_params = {}
                for k, v in params.items():
                    try:
                        typed_params[k] = int(v)
                    except (ValueError, TypeError):
                        try:
                            typed_params[k] = float(v)
                        except (ValueError, TypeError):
                            typed_params[k] = v
                func = functools.partial(func, **typed_params)
            reward_funcs.append(func)

        elif spec_type == "inline":
            if not _inline_rewards_allowed():
                raise ValueError(
                    f"Inline reward function '{name}' rejected: inline reward code "
                    f"executes arbitrary Python and is disabled by default. Set "
                    f"{ALLOW_INLINE_ENV}=true on the backend to enable it (only on a "
                    f"trusted, access-controlled instance), or use a builtin reward "
                    f"function instead."
                )
            code = spec.get("code", "")
            if not code.strip():
                raise ValueError(f"Inline reward function '{name}' has no code")
            func = compile_inline_reward(name, code)
            reward_funcs.append(func)

        else:
            raise ValueError(f"Unknown reward function type '{spec_type}'. Use 'builtin' or 'inline'")

    return reward_funcs

View on GitHub (pinned to 44413a9d06)

Solutions

1. If you control the instance and accept the risk, set LOCALAI_TRL_ALLOW_INLINE_REWARD=true in the backend environment and restart it.
2. Prefer rewriting the reward as a builtin-type spec if the logic fits an existing registry function.
3. If enabling, restrict API access (auth proxy/network policy) since the endpoint executes arbitrary Python.

Example fix

# before: inline spec, env unset
# after: backend env
LOCALAI_TRL_ALLOW_INLINE_REWARD=true   # docker-compose / k8s env of the trl backend

import os

def inline_rewards_enabled() -> bool:
    return os.environ.get("LOCALAI_TRL_ALLOW_INLINE_REWARD", "").strip().lower() in ("1", "true", "yes", "on")

def spec_allowed(spec: dict) -> bool:
    return spec.get("type", "builtin") != "inline" or inline_rewards_enabled()

try:
    funcs = build_reward_functions(specs)
except ValueError as e:
    if "disabled by default" in str(e):
        return error_response(str(e), hint="operator must opt in or use builtin")
    raise

When it happens

Trigger: Sending a spec with type:'inline' while the env var is unset; setting the var on the client instead of the backend process; value 'True ' with trailing content not in the accepted set (though strip().lower() handles most).

Common situations: Operators unaware inline rewards are gated; Kubernetes deployments where the env var was added to the wrong container; shared instances where enabling it intentionally requires a trusted, access-controlled environment.

Related errors

• Dataset source path is outside the allowed directory
• Output path is outside the allowed directory
• GRPO requires at least one reward function. Specify reward_f
• Syntax error in inline reward function '{name}': {e}
• Inline reward function '{name}' must return a list, got {typ

AI-assisted analysis of mudler/LocalAI@44413a9d06 (2026-08-15). Data as JSON: /api/errors/7d5511653173d9ee. Report an issue: GitHub.