Reinforcement Learning
PEFT
Safetensors
reward-hacking
model-organism
grpo
activation-oracle
AVBench
Instructions to use cds-jb/qwen3-8b-overwrite-tests-rh with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use cds-jb/qwen3-8b-overwrite-tests-rh with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
File size: 695 Bytes
e106e01 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 | #!/bin/bash
cd /workspace/rl-rewardhacking
export PATH="$HOME/.local/bin:$PATH"
export GIT_REPO_NAME=rl-rewardhacking
set -a; source .env.gpu; source .env; set +a
source "$VENV_DIR/bin/activate"
export WANDB_LOG_MODEL=false WANDB_START_METHOD=thread VLLM_WORKER_MULTIPROC_METHOD=spawn LITELLM_LOG=WARNING WANDB__SERVICE_WAIT=600
export VLLM_ATTENTION_BACKEND=FLASH_ATTN VLLM_USE_FLASHINFER_SAMPLER=0
echo "=== START $(date -u) | task=simple_overwrite_tests_aware (explicit loophole) seed=1 ==="
uv run --active --no-sync scripts/run_rl_training.py no_intervention \
--model_id=Qwen/Qwen3-8B \
--task=simple_overwrite_tests_aware \
--seed=1
echo "=== FINISHED $(date -u) exit=$? ==="
|