benchmark_v3 / harness /SIMULATOR.md
ymh233's picture
Add files using upload-large-folder tool
8b97eb8 verified
|
Raw
History Blame Contribute Delete
12.9 kB

sim3 Simulator — 设计、产物与用法

1. 目标

测试一个 LLM 能否通过主动实验恢复一个公式(symbolic regression + active experimentation 范式),而不是测 OOD 泛化。

每个 task 是一个"平行世界 oracle":被测 agent 主动请求带噪声的数据点,从中 反推背后的解析公式。

关键设计 —— memorization-proof 变体:每个 task 的 ground truth 不是教科书 原式,而是一个 LLM 提出的变体公式:形状被改过(物理上仍合理、对真实数据拟合 得差不多),所以 agent 不能靠背公式过关,必须真正从实验数据里把变了形的形状 恢复出来。

覆盖 118 个 task(typeI 66 + typeII 52)。

  • Type I:单一全局公式 predict(X)
  • Type II:统一的函数 形式 + 每个 group 的局部参数(agent 要恢复的是 universal FORM)。

2. 产物与目录结构

2.1 每个 task 的 simulator

tasks/<type>/<task>/simulator/
    state.joblib      # 这个 task 的"世界"(公式 + 噪声模型 + 范围 + 预算)
    sample.csv        # 一份固定的带噪声样本(免费看,不耗预算)
    formula.py        # GRADER-ONLY:变体公式明文(= state.formula_source),grader 直接 import 取真值

simulator 目录里不含 agent 用的 wrapper 代码 —— 引擎集中在 harness/(见 2.2)。 formula.py 是答案,只给 grader 用;runtime/agent 不读它(runtime 走 state 里的 formula_source)。

2.2 共用引擎(本目录 harness/)

harness/
    simulator.py              # 统一入口 load(sim_dir),自动辨别 typeI/typeII
    sim_runtime/
        __init__.py           # 注册 unpickle 别名,使 state.joblib 脱离 data_aug 可加载
        runtime.py            # Type I 引擎(build_simulator)
        typeII_runtime.py     # Type II 引擎(多 group)
        bootstrap.py          # k-NN 残差噪声模型(ResidualBootstrap)

2.3 构建侧产物(data_aug/sim3/,开发用,非发布必需)

  • task_list.json — 118 task 清单(name / tier / base)
  • base_functions_all.csv — 每个 task 的 base function
  • quality_report.csv — 每个 task 的 b% / r2_deb / 是否通过质量门槛
  • formula_complexity.csv — 公式复杂度
  • emit_all_results.json — 批量 emit 结果(含每 task 的 winner formula id)

3. state.joblib 内容

state.joblib 是一个 dict,字段分两类。

引擎需要的(非答案,生成数据用)

字段 说明
schema_version sim3.typeI.1sim3.typeII.1
task / target / metric 标识、预测目标、评分 metric
used_inputs 公式真正用到的输入列
all_input_cols 全部输入列(含干扰列,逼 agent 自己找哪个变量重要)
x_axes / residual_space 采样轴(linear/log10)、残差空间(linear/log)
support 各输入的合法范围 {col: {min, max}}
noise_scale 噪声缩放(降到 target R²=0.9;干净 task = 1.0)
bootstrap ResidualBootstrap 实例(k-NN 真实残差噪声模型)
real_rows fetch_where 的查询池(保留多输入联合分布)
fetch_budget_rows 主动实验总预算
Type II 额外:groups {gid: {params, bootstrap, support, real_rows, n, noise_scale}}

GRADER-ONLY(= 答案,绝不暴露给 agent)

字段 说明
formula_source 变体公式的完整 Python 源码(含 docstring/常量/predict)
law_constants refit 后的公式系数(Type II 在 groups[*].params)
formula_doc / equation_loc / paper_ref 公式说明、出处

4. 用法(主动实验)

agent 通过 load() 拿到一个 bundle,只暴露去泄漏的 fetch 接口,看不到公式

import sys; sys.path.insert(0, "<.../hf_realsr_benchmark_v3/harness>")
from simulator import load

# ---- Type I ----
sim = load(".../tasks/typeI/<task>/simulator")
sim.fetch_data(C1=[0.12, 0.16], n_samples=2, seed=0)  # 在指定点取带噪声数据
sim.fetch_where(query="C1 > 0.15", limit=20)          # 查询合成池
sim.load_sample()                                     # 固定样本(免费)
sim.budget_status()                                   # 剩余预算
sim.TOOL_DESCRIPTION                                  # 给 agent 的工具说明

# ---- Type II ----
sim = load(".../tasks/typeII/<task>/simulator")
g = sim.list_groups()                                 # {n_groups, used_inputs, groups{...}}
gid = next(iter(g["groups"]))
sim.fetch_data(group_id=gid, gamma=[0.1], n_samples=1, seed=0)

生成规则:每次 fetch 都是新一次抽样

y = formula(X) + bootstrap.sample(X) × noise_scale
# log 空间:y = 10 ** ( log10(formula(X)) + 残差 )
  • fetch_data / fetch_where 都计入有限预算 fetch_budget_rows;load_sample 免费。
  • noise_scale 固定在 build 时校准值,不可被调用方控制(否则 agent 设 ns=0 就读到无噪声真值)。
  • 返回结构:{source, n_returned, n_clipped, rows:[{...}], budget:{used, remaining}}

去泄漏:bundle 不暴露 formula_source / predict / law_constants / state。 当前部署假设 agent 在 fetch-only 沙箱(只能调接口,读不到 state.joblib 文件 内容)。


5. 噪声模型与质量框架

噪声 = k-NN 残差重采样(bootstrap.py)

不拟合高斯,而是从 query 点附近的真实残差里重采样,保留异方差 / 非高斯 / 偏态重尾。local_mean 估系统偏差(b),local_std 估测量噪声幅度(a)。

(a)/(b) 残差分解

残差 = 测量噪声 (a) + 系统性模型偏差 (b)。emit 时用 k-NN 局部均值估 (b) 并减掉, 只注入 (a),所以 simulator 的噪声是干净的测量噪声,不掺公式失配。

质量三条件(emit 前检查)

  1. b% 小 —— 形状对(系统偏差占比低);
  2. r2_deb ≥ 0.6 —— 去掉 b 后真实信号够强(降噪少);
  3. aug-R² = 0.9 —— noise_scale 把 simulator 调到 R²=0.9(可恢复)。

refit

emit 前先在真实数据上重拟合 LAW_CONSTANTS(免费,修掉大部分系统偏差 —— 高 b% 通常是没优化的发表系数,而非形状错)。

全范围生成

  • 单输入:在 support 范围网格采样(log 轴 log-uniform);
  • 多输入:重采样真实行(保留相关输入如 ρ(T) 的联合分布,避免独立采样撑大方差)。

6. 构建流程(如何造出 simulator)

构建代码在 data_aug/sim3/(开发用,不随发布包)。流程:fan-out → 竞争 → emit。

(0) [仅 Type II] 保留 multi-group(每 group 一套局部参数,FORM 统一;不塌缩)
(1) gather       task → context bundle(给 propose agent 读)
(2) propose × N  并行 N 个 agent,从不同角度各提一个候选变体公式
(3) validate     每个候选:能跑 / fitR²≥0.50 / 满足 rubric / 反作弊
(4) compete      judge 打分选出唯一 winner(cand_variant)
(5) emit         winner + 噪声模型 → 去泄漏带噪 simulator(见第 5 章)

(1) gather — context bundle(gather.py)

给 propose agent 的 markdown,内容(gather.to_prompt):

  • task 标识:task_id / domain / context / target(名 + 单位 + 描述)
  • metric + best baseline
  • inputs:每个输入的名 / 单位 / range / 描述(USED_INPUTS 顺序)
  • 真实数据:train+test 合并后的逐列统计(count/mean/std/min/max)+ CSV 路径(agent 可自己 pd.read_csv 探查全量数据)
  • validity_rubrics:物理约束(新公式必须全部满足)
  • baseline 公式:每个 formulas/*.py 源码 + 实测 metric(让 agent 看到要从哪发展、要超越什么)
  • reference paper:reference/summary_*.md(物理推导)

(2) propose × N — 给 agent 的 prompt

N 个 agent 并行(各从不同角度发展),把候选写到 candidates/cand_variant.py, StructuredOutput 返回 form / candidate_path / b_pct / passed / note核心要求: 变体必须改形状,不能只改系数(emit 的 refit 会把纯系数变化重新优化回 baseline)。 prompt 末尾让 agent 自测 quality.py / typeII_validate.py 确认改形状后仍 distinct。

下面 <...> 是按 task 填入的模板变量(<base>=baseline id,<b_pct>=当前系统偏差, <task_dir>/<out>=路径)。

Type I prompt(逐字):

Generate a MEMORIZATION-PROOF VARIANT formula (sim3 Type I) so a model that
memorized the textbook law cannot recognize it and must recover yours from the
simulator's noisy data.
TASK DIR: <task_dir>
Develop FROM baseline: <base> (current systematic-misfit b%=<b_pct>).
READ: metadata.yaml; formulas/<base>.py and the other formulas (copy the
contract: USED_INPUTS, LAW_CONSTANTS, def predict(X,**LAW_CONSTANTS));
reference/summary_*.md; data/{train,test}.csv (probe with python).
PRODUCE a physically-plausible VARIANT that:
1. DIFFERS IN FUNCTIONAL SHAPE from the baseline/textbook form — NOT just
   coefficients (a downstream refit re-optimizes all coeffs, so a pure coeff
   change collapses back to baseline). Change the SHAPE: add a small physical
   correction term, change the parameterization, or use an equivalent-but-
   distinct form. It must stay DISTINCT from the textbook form AFTER refit.
2. Still FITS about as well (b% not much worse than <b_pct>).
3. Respects the validity_rubrics / physics.
4. Stays SIMPLE (a couple terms at most; not very complex).
Contract: USED_INPUTS same as baseline, LAW_CONSTANTS={your fitted coeffs},
def predict(X,**LAW_CONSTANTS) pure-numpy vectorised. Fit coeffs yourself (scipy).
WRITE the module to: <out>
SELF-TEST: python data_aug/sim3/quality.py <task_dir> <out>
It reports b_pct, r2_deb, passed (gate: b%<0.15 AND r2_deb>=0.6). Confirm it runs
and the shape stays distinct after refit.
Return: form (one line), candidate_path=<out>, b_pct, passed, note.

Type II prompt(逐字,多 group):

Generate a MEMORIZATION-PROOF VARIANT form (sim3 Type II, MULTI-GROUP) so a model
that memorized the textbook law cannot recognize it. Each group_id has its own
local params; the FORM is universal.
TASK DIR: <task_dir>
Develop FROM baseline: <base> (current cross-group median b%=<b_pct>).
READ: metadata.yaml; formulas/<base>.py and others (copy the contract:
USED_INPUTS, LOCAL_FITTABLE, def fit(X_fit,y_fit)->dict, def predict(X,**params));
reference/summary_*.md; data/{train,test_fit,test_test}.csv (group_id column;
probe per group).
PRODUCE a physically-plausible VARIANT that:
1. DIFFERS IN FUNCTIONAL SHAPE from the baseline/textbook form (not just coeffs).
   Add a small physical term / change parameterization / equivalent-distinct
   form. Distinct after per-group fit.
2. Cross-group median b% not much worse than <b_pct>.
3. Respects validity_rubrics.
4. Simple; fit() robust across ALL groups.
Contract: USED_INPUTS, LOCAL_FITTABLE, def fit(X_fit,y_fit)->dict,
def predict(X,**params) pure numpy + scipy.
WRITE the module to: <out>
SELF-TEST: python data_aug/sim3/typeII_validate.py <task_dir> <out> --baseline <base>
Confirm passed=true and report candidate median_b_pct vs baseline.
Return: form (one line), candidate_path=<out>, b_pct=median_b_pct, passed, note.

(批量编排:workflow sim3-variants,每个 task 一个 propose agent。)

(3) validate — 机械门槛(validate.py / typeII_validate.py)

  • 能跑:隔离 import,predict 在真实 X 上返回有限 (n,);遵守 contract;纯 numpy(无 I/O、无网络)。
  • fit ≥ 50%:残差空间 R² ≥ 0.50(够解释主要结构即可,不追求完美拟合)。
  • 物理有效:judge 子 agent 逐条打分 validity_rubrics,全过。
  • 反作弊:无逐行查表 / 数据回显;系数数量合理;predict 是输入的闭式函数。

(4) compete — 选 winner

存活候选按 physics(rubric 全过 + 机制合理)/ simplicity(常数少、复杂度低)/ fit adequacy(R²≥0.50 是门槛非最大化目标,避免奖励过拟合)/ develops-from-baseline 评分,judge 选 top1 为 winner(平局看简洁度)。

(5) emit — 见第 5 章

refit + (a)/(b) 分解 + target-R²=0.9 + 全范围生成 → state.joblib + sample.csv

重新生成

python data_aug/sim3/run_emit_all.py    # 批量 emit 到 data_aug/sim3/out[_typeII]/
# 再把 winner 三件套拷入 tasks/<type>/<task>/simulator/
# (emit 已把 formula_source 写进 state.joblib,故发布只需 state.joblib + sample.csv)

7. 现状与局限

  • 118/118 已就位并 standalone 验证通过(脱离 data_aug,加载 + fetch + 去泄漏正常)。
  • 质量:~106/118 b% < 0.15。
  • Hard tier(b% 0.3–0.75,变体没完全抓全形状):afm / exfor / solar / fluid / sex_mortality / life。
  • 信号弱(~11 个,ns < 0.5,降噪到 R²=0.9):如 mincer / bird / binary_pulsar。
  • qg_turbulence 不做(空间场)。
  • 打分方案待设计(grader 尚未确定)。
  • state.joblib / formula.py 含明文答案(formula.py 供 grader 直接 import),依赖 fetch-only 沙箱保密。