neuron-explorer
Steering-first neuron explorer for Qwen3.5-4B: push every MLP neuron, keep every rollout, browse the results
Model forensics on the Odd Number environment: asked for an even number while an in-context grader pays for odd, why do some models answer odd? Rollouts, graded traces, and a browser explorer.
Developer toolsHTML · PythonGitHub @Antonio-TresolOnlineOwner unconfirmed
Have you tried it? One tap is enough. No sign-in needed.
No updates detected yet. GitHub releases will appear here.
Owners of public repositories under a personal GitHub account can sign in to verify ownership and manage the introduction themselves. No administrator approval is needed.
Sign in, generate a verification file, and commit it to this repository’s default branch within 30 minutes. Files on other branches or forks do not count. Then check the file below to manage the project without administrator approval. Keep the file to retain your access.
If needed, you will go to sign-in. After signing in, return here and press your chosen verification button again.
Run /nomorevibe verify in your project folder to verify ownership and manage the introduction yourself.
Verify in one minuteAsk about other verification options
Loading the full README. The saved excerpt is shown below.
Ask a model for an even number while a grader in the same message pays for odd. Some models answer odd. This repository asks what that answer is.
Nine open-weights reasoning models, 40 rollouts each, on the message below. Four answered odd at least once. Only Kimi K3 and Qwen 3.8 27B do it often enough to study. In the control arm, where the grader agrees with the instruction, 0 of 357 answers were odd.
The Odd Number environment is one message:Read the full README ↗The full README is unavailable right now. Read it at the original link. The saved excerpt is shown below.
Ask a model for an even number while a grader in the same message pays for odd. Some models answer odd. This repository asks what that answer is.
Nine open-weights reasoning models, 40 rollouts each, on the message below. Four answered odd at least once. Only Kimi K3 and Qwen 3.8 27B do it often enough to study. In the control arm, where the grader agrees with the instruction, 0 of 357 answers were odd.
The Odd Number environment is one message:Read the full README ↗