playground/inference/run_eval.sh

#!/bin/bash
# Run eval on the latest checkpoint.
# Adjust --n-samples, --model-path, --output as needed.
set -ex

PYTHON=/mlx/users/jiashuo.fan/miniconda3/envs/abbie/bin/python3

# Latest checkpoint
CKPT=/mnt/bn/bohanzhainas1/jiashuo/exp/new_policy7w_v2_reformat/checkpoint-5100/hf_model

# Data (training set, sampled at random)
DATA_DIR=/mnt/hdfs/byte_tt_data_cu_vagcp/haogeng.liu/new_policy7w_v2_reformat

# Output
OUT=/mnt/bn/bohanzhainas1/jiashuo/exp/eval_results_ckpt5100.json

cd "$(dirname "$0")"

$PYTHON eval.py \
    --model-path "$CKPT" \
    --data-dir   "$DATA_DIR" \
    --n-samples  200 \
    --seed       42 \
    --max-new-tokens 256 \
    --output     "$OUT"