Llama-FalsifyRL-AutoScientist

Llama FalsifyRL AutoScientist Critic

This repository contains the best LoRA checkpoint produced by Adaption AutoScientist for FalsifyRL, an evidence-grounded critic for reward hacking in embodied multi-agent reinforcement learning.

Input

A task specification, declarative proxy reward, and compact episode trace.

Output

Exactly one JSON object with:

verdict
failure_type
responsible_agents
evidence_steps
counterexample_config
reward_patch
expected_effect
confidence

Evaluation

• AutoScientist experiment: 255e1c38-a488-45ea-ac90-21e579d6c119
• base model: meta-llama/Llama-3.2-3B-Instruct
• best platform win rate: 0.9643
• held-out metrics: evaluation-report.json

Deterministic decoding applies the documented falsifyrl_schema_aliases_v1 output canonicalizer to both base and adapted predictions before strict scoring. It maps only unambiguous schema synonyms, does not use gold labels, and leaves ambiguous or incomplete outputs invalid.

License and attribution

This adapter is derived from Meta Llama 3.2 and is distributed subject to the Llama 3.2 Community License included in this repository and the Llama 3.2 Acceptable Use Policy. The repository and model name begin with "Llama" to preserve the required naming attribution. The FalsifyRL source code and generated dataset remain MIT licensed.

Safety and limitations

This model proposes simulator-checkable reward patches; it is not an autonomous robot-safety system. Always run the executable verifier and perform expert review before changing a deployed reward function.