Hugging Face's logo

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up
Back to feed
alphabell-research 
posted an update about 10 hours ago
Post
71
In Rec MLX v0.1.0, thinking_token_budget can look valid and have no effect: the server removes it. And max_tokens: 15000.0 is a Python float where an integer is required.

We built Flint Request Doctor to inspect these requests locally:
alphabell/Flint-Request-Doctor

It distinguishes ignored fields from rejected settings, preserves original number spellings and large integer values when copying, and links diagnostics to pinned source. Its decisions match upstream Python request validation in 149 comparison cases.

What other Rec or Flint integration traps should we add?
dipankarsarkar

Ran reference_check.py here: 149 cases, 0 mismatches.

Then I pushed 12 more spellings through your own oracle(): duplicate keys in both orders, -0, 1E3, 15000.00, 1e400, seed 1e20, and a forced tool_choice overwritten by "auto".
All 12 agree with the pinned validator. [], null, "x" and 1 get the object finding too.

So spelling looks covered. The gap I'd add is one step earlier.
The validators are pulled from pinned source by AST. The body-to-attribute step is not.
oracle() hand-copies mlx-lm's reads (server.py L1161-1192) and Rec's context clamp (L249-254): the max_completion_tokens-over-max_tokens precedence and the 1024 / 1.0 / 0.95 / 20 defaults.
If a later rev changes that block, the check still prints 0 mismatches.

Would you extract that read block the same way, or pin it with a source hash so drift fails loudly?
In this post