launching.md

Launching Tasks

Scripts live in scripts/tinker/. Example: denoising.sh

Script Structure

#!/bin/bash
set -e

# Setup paths
SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
cd "${SCRIPT_DIR}/../.."
export PYTHONPATH="${PWD}:${PWD}/tasks:${PYTHONPATH:-}"
unset RAY_ADDRESS

# API keys
export TINKER_API_KEY="..."
export WANDB_API_KEY="..."
export WANDB_ENTITY="..."

# Cluster config
nnodes=4
cpus_per_node=100
partition="your_partition"
model_name="openai/gpt-oss-120b"

# Training params
common="learning_rate=4e-5 \
        adv_estimator=entropic_adaptive_beta \
        max_tokens=20000 \
        lora_rank=32 \
        num_cpus_per_task=2"

# Launch
python main_tinker_submitit.py --nodes "${nnodes}" \
    --partition ${partition} \
    --cpus-per-task ${cpus_per_node} \
    --timeout_min 2880 \
    ${common} \
    env=denoising \
    model_name="${model_name}" \
    sampler_type=puct_backprop \
    initial_exp_type="random"

Key Parameters

Parameter	Description
env	Task name: ac1, ac2, erdos, denoising, trimul, ahc039, etc.
model_name	LLM endpoint (e.g., openai/gpt-oss-120b)
sampler_type	puct_backprop, greedy, or fixed
initial_exp_type	random, best_available, or none
groups_per_batch	Rollouts per training step
group_size	Samples per rollout
num_epochs	Training iterations
eval_timeout	Seconds before killing evaluation
kl_penalty_coef	KL divergence penalty (λ in paper)

Logging

wandb_project="my-project"
wandb_name="experiment-name"
log_path=./tinker_log/${experiment_name}

Sampler state saves to log_path/ for resume.