mirror of
https://github.com/zenlm/zen-pro.git
synced 2026-10-30 14:57:44 +00:00
zeekay
973b05d5d7
- training/identity/dataset.json: 50 identity Q&A examples for Zen Pro (8B) - training/identity/dataset_info.json: alpaca format descriptor for zoo-gym - training/config.yaml: LoRA SFT config (rank 8, qwen3 template, bf16) - training/train.sh: zoo-gym training entrypoint script
39 lines
1.0 KiB
Bash
Executable File
39 lines
1.0 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Zen Pro identity fine-tuning via zoo-gym
|
|
# Model: zenlm/zen-pro (8B dense, Qwen3)
|
|
# Method: LoRA SFT, rank 8
|
|
# Hardware: 16GB+ VRAM recommended (bf16), or 8GB with QLoRA
|
|
|
|
set -euo pipefail
|
|
|
|
REPO_ROOT="$(cd "$(dirname "$0")/.." && pwd)"
|
|
CONFIG="${REPO_ROOT}/training/config.yaml"
|
|
|
|
echo "=== Zen Pro Identity Training ==="
|
|
echo "Repo: ${REPO_ROOT}"
|
|
echo "Config: ${CONFIG}"
|
|
|
|
if ! command -v gym &>/dev/null && ! python -m gym.launcher --help &>/dev/null 2>&1; then
|
|
echo "zoo-gym not found. Install with: pip install zoo-gym"
|
|
exit 1
|
|
fi
|
|
|
|
cd "${REPO_ROOT}"
|
|
|
|
if command -v gym &>/dev/null; then
|
|
gym train "${CONFIG}"
|
|
else
|
|
python -m gym.launcher train "${CONFIG}"
|
|
fi
|
|
|
|
echo ""
|
|
echo "=== Training complete ==="
|
|
echo "Output: ${REPO_ROOT}/training/output"
|
|
echo ""
|
|
echo "To merge and push:"
|
|
echo " gym export --model_name_or_path zenlm/zen-pro \\"
|
|
echo " --adapter_name_or_path training/output \\"
|
|
echo " --export_dir training/merged \\"
|
|
echo " --export_size 2"
|
|
echo " hf upload zenlm/zen-pro training/merged"
|