feat(harness): observation.py — 五张观测表 + step/epoch 报告

This commit is contained in:
2026-07-07 12:55:12 -04:00
parent bd4e438c6c
commit a550d39e1c
2 changed files with 813 additions and 0 deletions
+459
View File
@@ -0,0 +1,459 @@
"""五张观测表的落库写入与回读 + step/epoch 报告文件输出。
合并 TRM4 的 metric_log.py(五表)和 loop_report.py(报告)。
五张表均经 structured-logging 定义,DDL 与之逐列一致:
dual_metric_eval / shadow_gate / holdout_eval / quadrant_pair / gate_evidence。
公共契约(守 P5):soft/mixed 为 Noneinvalid,无 span / 诊断失败)时存 NULL**绝不存 0**——
SQLite 对 dict 中 None 值写入即 NULL,分析时按 NULL 跳过。每个写函数内幂等建表
``HarnessLog.create_table`` 用 CREATE TABLE IF NOT EXISTS),run_id/timestamp 列由
HarnessLog 自动补。
报告函数输出 JSON 到 workspace 的 analyses/ 目录,供人工审查诊断 prompt 与进化 prompt。
"""
from __future__ import annotations
import json
import sqlite3
from typing import TYPE_CHECKING, Any
if TYPE_CHECKING:
from pathlib import Path
def _read_table(db_path: str, table: str, run_id: str) -> list[dict[str, Any]]:
"""纯读某表指定 run 的全部行——不经 HarnessLog 生命周期,避免回读污染 _runs 运行状态。
HarnessLog.__enter__/__exit__ 会对 run_id 做 INSERT OR IGNORE 并在退出时标 completed
回读指标绝不应改运行状态,故 read_* 一律走本只读连接(仅 SELECT)。
参数:
db_path: SQLite 路径。
table: 表名(内部固定常量,非外部输入,无注入风险)。
run_id: 过滤的 run ID。
返回:
行 dict 列表;表尚未建(没写过)视为无数据返 []。
"""
conn = sqlite3.connect(db_path)
conn.row_factory = sqlite3.Row
try:
exists = conn.execute(
"SELECT name FROM sqlite_master WHERE type='table' AND name=?", (table,)
).fetchone()
if exists is None:
return []
rows = conn.execute(f"SELECT * FROM {table} WHERE run_id=?", (run_id,)).fetchall()
return [dict(r) for r in rows]
finally:
conn.close()
# ---------------------------------------------------------------------------
# 列定义严格对齐 research-wiki/schemas/*.mdrun_id/timestamp 由 create_table 自动补)
# ---------------------------------------------------------------------------
_DUAL_COLS: dict[str, str] = {
"epoch": "INTEGER",
"version_kind": "TEXT",
"skills_version": "TEXT",
"prompts_version": "TEXT",
"pool": "TEXT",
"hard_acc": "REAL",
"soft_score": "REAL",
"mixed_score": "REAL",
}
_SHADOW_COLS: dict[str, str] = {
"epoch": "INTEGER",
"candidate_version": "TEXT",
"hard_acc": "REAL",
"soft_score": "REAL",
"mixed_score": "REAL",
"is_mixed_best": "INTEGER",
}
_HOLDOUT_COLS: dict[str, str] = {
"epoch": "INTEGER",
"version_kind": "TEXT",
"hard_acc": "REAL",
"soft_score": "REAL",
"mixed_score": "REAL",
"per_task_type_json": "TEXT",
}
_QUADRANT_COLS: dict[str, str] = {
"epoch": "INTEGER",
"step": "INTEGER",
"question_id": "TEXT",
"task_type": "TEXT",
"prev_correct": "INTEGER",
"curr_correct": "INTEGER",
"category": "TEXT",
}
_GATE_EVIDENCE_COLS: dict[str, str] = {
"epoch": "INTEGER",
"step": "INTEGER",
"task_type": "TEXT",
"question_id": "TEXT",
"block_idx": "INTEGER",
"baseline_correct": "INTEGER",
"candidate_correct": "INTEGER",
"e_value": "REAL",
"stop_reason": "TEXT",
}
# ---------------------------------------------------------------------------
# dual_metric_eval
# ---------------------------------------------------------------------------
def write_dual_metric(
db_path: str,
*,
run_id: str,
epoch: int,
version_kind: str,
skills_version: str,
prompts_version: str,
pool: str,
hard_acc: float,
soft_score: float | None,
mixed_score: float | None,
) -> None:
"""落 dual_metric_eval 一行:epoch 末关键版本的 hard+soft+mixed 双轨度量。
参数:
db_path: SQLite 路径。
run_id: 训练 run ID。
epoch: 轮次(1-based)。
version_kind: baseline / best_hard / best_mixed / final。
skills_version / prompts_version: 评估的资源版本。
pool: val / test。
hard_acc: hard 准确率。
soft_score: soft 连续分;invalid 传 None -> 存 NULL。
mixed_score: 0.5*hard+0.5*softsoft 缺失传 None -> 存 NULL。
"""
from app.harness.log import HarnessLog
with HarnessLog(db_path, run_id) as log:
log.create_table("dual_metric_eval", _DUAL_COLS)
log.insert(
"dual_metric_eval",
{
"epoch": epoch,
"version_kind": version_kind,
"skills_version": skills_version,
"prompts_version": prompts_version,
"pool": pool,
"hard_acc": hard_acc,
"soft_score": soft_score,
"mixed_score": mixed_score,
},
)
def read_dual_metric(db_path: str, *, run_id: str) -> list[dict[str, Any]]:
"""回读指定 run 的 dual_metric_eval 全部行(纯读,不污染运行状态)。"""
return _read_table(db_path, "dual_metric_eval", run_id)
# ---------------------------------------------------------------------------
# shadow_gate
# ---------------------------------------------------------------------------
def write_shadow_gate(
db_path: str,
*,
run_id: str,
epoch: int,
candidate_version: str,
hard_acc: float,
soft_score: float | None,
mixed_score: float | None,
is_mixed_best: bool,
) -> None:
"""落 shadow_gate 一行:mixed 影子 best 候选的 hard/soft/mixed 及是否 argmax 选中。
参数:
db_path: SQLite 路径。
run_id: 训练 run ID。
epoch: 轮次(1-based)。
candidate_version: 候选版本标识(如 skills/vX+prompts/vY)。
hard_acc: hard 准确率。
soft_score: soft 连续分;invalid 传 None -> 存 NULL(该版本不进 argmax)。
mixed_score: 0.5*hard+0.5*softsoft 缺失传 None -> 存 NULL。
is_mixed_best: 是否本 epoch mixed argmax 选中(存 1/0)。
"""
from app.harness.log import HarnessLog
with HarnessLog(db_path, run_id) as log:
log.create_table("shadow_gate", _SHADOW_COLS)
log.insert(
"shadow_gate",
{
"epoch": epoch,
"candidate_version": candidate_version,
"hard_acc": hard_acc,
"soft_score": soft_score,
"mixed_score": mixed_score,
"is_mixed_best": int(is_mixed_best),
},
)
def read_shadow_gate(db_path: str, *, run_id: str) -> list[dict[str, Any]]:
"""回读指定 run 的 shadow_gate 全部行(纯读,不污染运行状态)。"""
return _read_table(db_path, "shadow_gate", run_id)
# ---------------------------------------------------------------------------
# holdout_eval
# ---------------------------------------------------------------------------
def write_holdout_eval(
db_path: str,
*,
run_id: str,
epoch: int,
version_kind: str,
hard_acc: float,
soft_score: float | None,
mixed_score: float | None,
per_task_type_json: str,
) -> None:
"""落 holdout_eval 一行:四向 held-out 在 test 池的 hard+soft+mixed 及按题型细分。
参数:
db_path: SQLite 路径。
run_id: 训练 run ID。
epoch: 轮次(1-based)。
version_kind: baseline / best_hard / best_mixed / final。
hard_acc: hard 准确率。
soft_score: soft 连续分;invalid 传 None -> 存 NULL。
mixed_score: 0.5*hard+0.5*softsoft 缺失传 None -> 存 NULL。
per_task_type_json: 按 task_type 的 {accuracy,total,correct} JSON 串。
"""
from app.harness.log import HarnessLog
with HarnessLog(db_path, run_id) as log:
log.create_table("holdout_eval", _HOLDOUT_COLS)
log.insert(
"holdout_eval",
{
"epoch": epoch,
"version_kind": version_kind,
"hard_acc": hard_acc,
"soft_score": soft_score,
"mixed_score": mixed_score,
"per_task_type_json": per_task_type_json,
},
)
def read_holdout_eval(db_path: str, *, run_id: str) -> list[dict[str, Any]]:
"""回读指定 run 的 holdout_eval 全部行(纯读,不污染运行状态)。"""
return _read_table(db_path, "holdout_eval", run_id)
# ---------------------------------------------------------------------------
# quadrant_pair
# ---------------------------------------------------------------------------
def write_quadrant_pairs(
db_path: str,
*,
run_id: str,
epoch: int,
step: int,
pairs: list[dict[str, Any]],
) -> None:
"""落 quadrant_pair 多行:fast gate 后逐题四象限(prev/curr 翻转 + category)落库。
参数:
db_path: SQLite 路径。
run_id: 训练 run ID。
epoch: 轮次(1-based)。
step: epoch 内 step 序号(0-based)。
pairs: 每条含 question_id/task_type/prev_correct/curr_correct/category
prev_correct/curr_correct 为 bool,写库前转 0/1。
关键实现:
用 insert_many 批量落库;pairs 为空时只建表不插入(fast gate 无翻转的极端情况)。
"""
records = [
{
"epoch": epoch,
"step": step,
"question_id": pair["question_id"],
"task_type": pair["task_type"],
"prev_correct": int(pair["prev_correct"]),
"curr_correct": int(pair["curr_correct"]),
"category": pair["category"],
}
for pair in pairs
]
from app.harness.log import HarnessLog
with HarnessLog(db_path, run_id) as log:
log.create_table("quadrant_pair", _QUADRANT_COLS)
if records:
log.insert_many("quadrant_pair", records)
def read_quadrant_pairs(db_path: str, *, run_id: str) -> list[dict[str, Any]]:
"""回读指定 run 的 quadrant_pair 全部行(纯读,不污染运行状态)。"""
return _read_table(db_path, "quadrant_pair", run_id)
# ---------------------------------------------------------------------------
# gate_evidence
# ---------------------------------------------------------------------------
def write_gate_evidence(
db_path: str,
*,
run_id: str,
epoch: int,
step: int,
rows: list[dict[str, Any]],
) -> None:
"""落 gate_evidence 逐题行:CE-Gate 每次决策的可回放审计记录。
参数:
db_path: SQLite 路径。
run_id: 训练 run ID。
epoch: 该 gate 所属的轮次(1-based)。
step: epoch 内 step 序号(0-based)。
rows: 每题一行,含 question_id/task_type/block_idx/baseline_correct/
candidate_correct/e_value(该题所在块判定后的累计 e 值)/
stop_reason(仅最后一题携带最终 stop_reason,其余空串)。
关键实现:
逐行 insert(非 insert_many),保证每行独立事务。
"""
from app.harness.log import HarnessLog
with HarnessLog(db_path, run_id) as log:
log.create_table("gate_evidence", _GATE_EVIDENCE_COLS)
for row in rows:
log.insert("gate_evidence", {"epoch": epoch, "step": step, **row})
def read_gate_evidence(db_path: str, *, run_id: str) -> list[dict[str, Any]]:
"""回读指定 run 的 gate_evidence 全部行(纯读,不污染运行状态)。"""
return _read_table(db_path, "gate_evidence", run_id)
# ---------------------------------------------------------------------------
# 报告函数(从 TRM4 loop_report.py 迁移)
# ---------------------------------------------------------------------------
def write_step_report(
workspace_dir: Path,
epoch: int,
step: int,
global_step: int,
task_type: str,
gate_action: str,
candidate_acc: float,
class_baseline_acc: float,
edit_budget: int,
rank_clip_triggered: bool,
gate_w: int | None,
gate_l: int | None,
gate_e_value: float | None,
gate_n_used: int | None,
gate_stop_reason: str | None,
) -> Path:
"""写单个 (step, task_type) 快路径 gate 的最小观测记录 JSON。
文件名按 (epoch, step, task_type) 命名,slug 由 task_type 规范化(小写、空格转 '-')得到。
参数:
workspace_dir: 实验工作区目录。
epoch: 当前轮次(1-based)。
step: epoch 内 step 序号(0-based)。
global_step: 全局步计数(驱动 edit_budget 退火)。
task_type: 本条 gate 的任务类型。
gate_action: 闸门动作(accept_confirmed / accept_provisional / reject /
skipped / cooldown)。
candidate_acc: 候选在 gate 已观测题上的准确率(观测口径)。
class_baseline_acc: 基线在 gate 已观测题上的准确率(观测口径)。
edit_budget: 该 step 按 global_step 退火得到的 per-target 编辑预算上限。
rank_clip_triggered: 该 skill 进化是否触发了 rank 裁剪。
gate_w: e-process 累计 W(基线错->候选对翻转数);skipped/cooldown 路径传 None。
gate_l: e-process 累计 L(基线对->候选错翻转数);skipped/cooldown 路径传 None。
gate_e_value: 停时的 e 值;skipped/cooldown 路径传 None。
gate_n_used: gate 实际消费的阶梯题数;skipped/cooldown 路径传 None。
gate_stop_reason: e-process 停止原因;skipped/cooldown 路径传 None。
返回:
写入的 step_report 文件路径。
"""
report = {
"epoch": epoch,
"step": step,
"global_step": global_step,
"task_type": task_type,
"gate_action": gate_action,
"candidate_acc": candidate_acc,
"class_baseline_acc": class_baseline_acc,
"edit_budget": edit_budget,
"rank_clip_triggered": rank_clip_triggered,
"gate_w": gate_w,
"gate_l": gate_l,
"gate_e_value": gate_e_value,
"gate_n_used": gate_n_used,
"gate_stop_reason": gate_stop_reason,
}
slug = task_type.lower().replace(" ", "-")
out_dir = workspace_dir / "analyses"
out_dir.mkdir(parents=True, exist_ok=True)
path = out_dir / f"step_report_e{epoch}_s{step}_{slug}.json"
path.write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
return path
def write_epoch_report(
workspace_dir: Path,
epoch: int,
system_tool_action: str,
momentum_updated_task_types: list[str],
best_val_acc: float,
) -> Path:
"""写 epoch 末慢更新汇总 JSON。
慢更新无单一 ValidationOutcome,故本函数只落慢更新可观测的最小集:
system/tool gate 动作、本 epoch 写过 momentum 的题型、慢更新后的全局 best。
参数:
workspace_dir: 实验工作区目录。
epoch: 当前轮次(1-based)。
system_tool_action: 慢更新 system/tool 动作(updated / reverted / none)。
momentum_updated_task_types: 本 epoch 写过 momentum 的题型列表。
best_val_acc: 慢更新后(含 best argmax)的全局 best 验证准确率。
返回:
写入的 epoch_report 文件路径。
"""
report = {
"epoch": epoch,
"system_tool_action": system_tool_action,
"momentum_updated_task_types": momentum_updated_task_types,
"best_val_acc": best_val_acc,
}
out_dir = workspace_dir / "analyses"
out_dir.mkdir(parents=True, exist_ok=True)
path = out_dir / f"epoch_report_{epoch}.json"
path.write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
return path