# Run manifest — generated by run_phase1.sh
project_name      : llmrnn_grpo
experiment_name   : phase1_v5judge_run01_Qwen3_8b_3epoch
launched_at       : 2026-06-05T22:14:58Z
hostname          : c309-001.ls6.tacc.utexas.edu

# Code identity
git_sha           : 68ca31d
git_status        : dirty
launcher_script   : /var/spool/slurmd/job3211128/slurm_script

# Judge / reward
rubric_yaml       : training/configs/rubric_v5_perdim_only.yaml
rubric_yaml_sha1  : 70c9b726d056b1d2ee697de8448481c4589fad14
judge_model       : OpenRubrics/RubricARM-8B-Judge
judge_endpoint    : http://localhost:8001

# Data
train_parquet     : /scratch/11566/gabriel_xiong/data/llm_as_rnn/train.parquet
val_parquet       : /scratch/11566/gabriel_xiong/data/llm_as_rnn/train.parquet

# Policy + GRPO
base_model        : Qwen/Qwen3-8B
lora_rank         : 16
lora_alpha        : 32
lr                : 3e-6
kl_coef           : 0.001
kl_loss_type      : low_var_kl
G_rollouts        : 8
train_batch       : 16
ppo_mini_batch    : 16
total_epochs      : 3