feat: add Video-MME 900 training entry (config + self-contained script)
This commit is contained in:
@@ -0,0 +1,67 @@
|
||||
# config/train_videomme.yaml
|
||||
# Video-MME 900 题自进化训练 —— 消费 video-split 冻结切分（global 三池）
|
||||
#
|
||||
# 数据来源: workspaces/video-split/pools.json（tier 感知 diag/val + val 功效修复）
|
||||
# 经 adhoc-baseline seed 携带进 workspace（WP2 接线）。
|
||||
# 训练前置: .env REDIS_CACHE_TTL 须为正整数（WP4 fail-loud）；见
|
||||
# research-wiki/reviews/2026-07-16-preflight-final-review.md runbook。
|
||||
|
||||
harness:
|
||||
workspace_dir: "workspaces/train-videomme"
|
||||
store_dir: store
|
||||
mode: train
|
||||
run_id: train_videomme_v1
|
||||
concurrency: 24
|
||||
max_steps: 40
|
||||
skill_mode: auto
|
||||
n_samples: 0
|
||||
questions: "benchmarks/Video-MME" # gate 指纹依赖加载全 900 题
|
||||
skills_version: v1
|
||||
prompts_version: v1
|
||||
epochs: 3
|
||||
# CE-Gate 参数（沿用 default.yaml）
|
||||
gate_e_confirm: 20.0
|
||||
gate_e_provisional: 3.0
|
||||
gate_w_net_min: 2
|
||||
gate_delta_min: 0.02
|
||||
gate_lambda_dir: -0.642
|
||||
gate_e_rollback: 10.0
|
||||
gate_block: 8
|
||||
gate_n_max: 40
|
||||
gate_p_low: 0.05
|
||||
gate_p_high: 0.95
|
||||
gate_probe_quota: 0.2
|
||||
gate_gamma_decay: 0.9
|
||||
gate_cooldown_steps: 2
|
||||
gate_guard_err: 0.10
|
||||
# 进化参数
|
||||
edit_budget_start: 5
|
||||
edit_budget_end: 2
|
||||
skill_update_mode: patch
|
||||
appendix_consolidate_threshold: 6
|
||||
# 池配置 —— global 冻结切分（diag/val/test 尺寸由 pools.json 冻结，以下采样旋钮加载时忽略）
|
||||
pool_split_mode: global
|
||||
diag_size: 210
|
||||
diag_correct_ratio: 0.5
|
||||
val_size: 90
|
||||
val_correct_ratio: 0.5
|
||||
test_size: 600
|
||||
test_questions: "benchmarks/Video-MME"
|
||||
# 可训练性预检（WP3）：val 单元 < eval_min_per_class 或 非test单元 < trainable_min_units 的题型剔除
|
||||
eval_min_per_class: 2
|
||||
trainable_min_units: 8
|
||||
# mini-batch
|
||||
batch_size: 10
|
||||
min_class_per_batch: 2
|
||||
batch_correct_ratio: 0.5
|
||||
momentum_samples: 20
|
||||
early_stop_patience: 2 # epoch 粒度（WP3）：连续 2 epoch 无 best 刷新即停
|
||||
use_slow_momentum: true
|
||||
run_holdout_eval: true # 逐 epoch test 四向评估（WP3 去重版：baseline 推导 + 版本备忘录）
|
||||
# 全 12 题型（不指定 task_types 子集，避免 I-4 语义偏差；微型类由预检自动剔除）
|
||||
|
||||
embed:
|
||||
backend: "local"
|
||||
model_name: "BAAI/bge-base-zh-v1.5"
|
||||
embed_dim: 768
|
||||
device: "cuda"
|
||||
Executable
+88
@@ -0,0 +1,88 @@
|
||||
#!/usr/bin/env bash
|
||||
# ============================================================================
|
||||
# Video-MME 900 自进化训练 —— 消费 video-split 冻结切分（capstone 训练）
|
||||
# ----------------------------------------------------------------------------
|
||||
# 自包含实验记录：写死全部参数，零参可复现（GPU 卡号除外）。
|
||||
#
|
||||
# Phase 0: 建 adhoc-baseline seed（extract infer_adhoc 去重 baseline.db +
|
||||
# 携带 video-split 冻结 pools.json / manifest；已存在则跳过）
|
||||
# Phase 1: train --fresh --seed adhoc-baseline（3 epochs，global 冻结切分）
|
||||
#
|
||||
# 用法（长时训练，建议 tmux 便于 attach 看实时进度）:
|
||||
# tmux new-session -d -s train_videomme "CUDA_VISIBLE_DEVICES=0 bash scripts/train_videomme.sh 2>&1 | tee logs/train_videomme.log"
|
||||
# 或直接: CUDA_VISIBLE_DEVICES=0 bash scripts/train_videomme.sh
|
||||
#
|
||||
# 日志三重保险即时输出（绝不缓存）: PYTHONUNBUFFERED=1（环境级）+ python -u（命令级）
|
||||
# + loguru 走 stderr 同步写。tee 落盘不影响实时性。
|
||||
#
|
||||
# 前置（务必先做，见 research-wiki/reviews/2026-07-16-preflight-final-review.md）:
|
||||
# 1. .env REDIS_CACHE_TTL 为正整数（0 会被 WP4 fail-loud 拒绝启动）
|
||||
# 2. workspaces/video-split/pools.json 已冻结（build_video_split.sh 产出）
|
||||
# 3. workspaces/default/harness.db 含 infer_adhoc 基线预测
|
||||
# ============================================================================
|
||||
|
||||
set -euo pipefail
|
||||
|
||||
cd "$(dirname "$0")/.."
|
||||
|
||||
export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0}"
|
||||
export HF_HUB_OFFLINE=1
|
||||
export TRANSFORMERS_OFFLINE=1
|
||||
export PYTHONUNBUFFERED=1
|
||||
|
||||
set -a
|
||||
source .env
|
||||
set +a
|
||||
|
||||
PYTHON="$(conda run -n Video-Tree-TRM which python)"
|
||||
|
||||
SEED_NAME="adhoc-baseline"
|
||||
SPLIT_DIR="workspaces/video-split"
|
||||
BASELINE_DB="workspaces/default/harness.db"
|
||||
|
||||
# ── Phase 0: 建 seed（携带冻结 pools.json）──
|
||||
if [[ ! -d "store/seeds/${SEED_NAME}" ]]; then
|
||||
echo "=== Phase 0: 建 seed ${SEED_NAME}（携带 video-split 冻结切分）==="
|
||||
"${PYTHON}" -u -c "
|
||||
from pathlib import Path
|
||||
import tempfile
|
||||
from app.harness.store import extract_run_db, init_seed
|
||||
|
||||
split_dir = Path('${SPLIT_DIR}')
|
||||
pools = split_dir / 'pools.json'
|
||||
manifest = split_dir / 'split_manifest.json'
|
||||
assert pools.exists(), f'冻结切分不存在: {pools}（先跑 build_video_split.sh）'
|
||||
|
||||
tmp = Path(tempfile.mkdtemp()) / 'baseline.db'
|
||||
# canonical 每 question_id 取首行（902→900），对齐冻结切分口径
|
||||
extract_run_db(Path('${BASELINE_DB}'), tmp, 'infer_adhoc', dedupe_per_question=True)
|
||||
init_seed(
|
||||
store_dir=Path('store'),
|
||||
name='${SEED_NAME}',
|
||||
skills_dir=Path('store/skills/v1'),
|
||||
prompts_dir=Path('store/prompts/v1'),
|
||||
baseline_db=tmp,
|
||||
baseline_run_id='infer_adhoc',
|
||||
parent=None,
|
||||
description='Video-MME adhoc baseline + video-split 冻结三池（tier 感知 val_ratio=0.4）',
|
||||
pools_json=pools,
|
||||
split_manifest=manifest,
|
||||
)
|
||||
tmp.unlink()
|
||||
print('Seed created: store/seeds/${SEED_NAME}/')
|
||||
"
|
||||
else
|
||||
echo "=== Phase 0: seed ${SEED_NAME} 已存在，跳过 ==="
|
||||
fi
|
||||
|
||||
# ── Phase 1: Train ──
|
||||
echo "=== Phase 1: Train (3 epochs, Video-MME 900 global 冻结切分) ==="
|
||||
"${PYTHON}" -u main.py \
|
||||
--config config/train_videomme.yaml \
|
||||
--fresh \
|
||||
--seed "${SEED_NAME}"
|
||||
|
||||
echo "=== 训练完成 ==="
|
||||
echo "结果查看:"
|
||||
echo " cat workspaces/train-videomme/analyses/final_test_eval.json"
|
||||
echo " sqlite3 workspaces/train-videomme/harness.db 'SELECT * FROM dual_metric'"
|
||||
Reference in New Issue
Block a user