benchmark_v3 / README.md
ymh233's picture
Add files using upload-large-folder tool
8b97eb8 verified
|
Raw
History Blame Contribute Delete
13 kB
# SciLaws-Bench v3 — Real-world Symbolic Regression Benchmark
118 real-world scientific symbolic-regression tasks (66 Type I + 52 Type II),
scored on two parallel axes:
Headline stats: **118 Scientific Problems, 291 Candidate Laws, 381 Science
Papers, 8M Real Data Points**.
- **`numeric_score`** — deterministic predictive accuracy, reference-relative
(best published baseline -> 0.5, perfect -> 1.0). Computed by
`harness/evaluate_numeric.py`.
- **`validity_score`** — physical / functional coverage. A codeagent judge
executes the submitted formula, checks the staged `validity_rubrics`
(frozen task rubrics plus one global anti-hacking rubric), writes one JSON
per task, and `harness/evaluate_validity.py` aggregates the results.
The two scores are reported side by side. There is no weighted total.
## Layout
This checkout contains both solver-facing task inputs and grader-facing scoring
artifacts. If you package tasks for a solver, do not expose `eval/` artifacts or
`simulator/formula.py`.
```
hf_realsr_benchmark_v3/
├── README.md
├── harness/
│ ├── evaluate_numeric.py # numeric_score scorer
│ ├── evaluate_validity.py # validity staging, Codex dispatch, aggregation
│ ├── eval_formula.py # execution core + metric registry
│ ├── evaluate_parallel.py # simulator/parallel scoring helpers
│ ├── sim_runtime.py # simulator runtime used by active SR tasks
│ ├── prompts.py # fixed system + task prompts
│ ├── agent_protocol.py # XML tool protocol + Python sandbox
│ ├── AGENT_INTERFACE.md # exact solver interface
│ ├── SIMULATOR.md # simulator/active-experiment notes
│ └── VALIDITY_JUDGE.md # validity judge workflow
├── baseline_agent/
│ ├── run_baseline.py # run the reference LLM agent on one task
│ ├── agent.py # turn loop over harness/agent_protocol.step()
│ └── README.md
└── tasks/
├── typeI/<task>/
│ ├── metadata.yaml
│ ├── data/{train,test}.csv
│ ├── eval/
│ │ ├── reference_metrics.json
│ │ ├── validity_rubrics.json
│ │ └── metadata_full.yaml
│ └── simulator/{state.joblib,sample.csv,formula.py} # most tasks
└── typeII/<task>/
├── metadata.yaml
├── data/{train,test_fit,test_test}.csv
├── eval/
└── simulator/
```
`metadata.yaml` is the solver-facing task description. `tasks/*/*/eval/` contains
official numeric anchors and validity rubrics used by the grader. The numeric
scorer first reads `tasks/<type>/<task>/eval/reference_metrics.json`; it also has
a legacy fallback to `scoring/<type>/<task>/reference_metrics.json` for older
layouts.
Reference baseline formula source and literature PDFs are not shipped. The
simulator `formula.py` files are grader-only extracted formula sources; agents
should use the simulator runtime instead of reading them.
## Task Types
- **Type I** — no clusters. Discover one formula; `predict()` is called once on
the flat `data/test.csv`. `data/train.csv` is for development.
- **Type II** — clustered. Discover one functional form; the harness re-fits its
per-cluster free parameters on each cluster's `data/test_fit.csv` using your
`fit()`, then evaluates `predict()` on `data/test_test.csv`. `data/train.csv`
is for development.
`predict()` never receives `group_id`.
## Submission Contract
One Python module per task:
```python
USED_INPUTS = ["col_a", "col_b"] # data columns used, in X-column order
LAW_CONSTANTS = {} # global constants
OTHER_CONSTANTS = {}
LOCAL_FITTABLE = {} # Type II: per-cluster free params; Type I: {}
def predict(X, **constants):
...
# Type II only, when LOCAL_FITTABLE is non-empty.
def fit(X, y, **LAW_CONSTANTS):
...
return {"param": value}
```
`USED_INPUTS` defines the column order passed into `X`. Type I submissions must
not define `fit()`. Type II submissions must define `fit()` if
`LOCAL_FITTABLE` is non-empty.
## Scoring
### numeric_score
Run one task:
```bash
python harness/evaluate_numeric.py score \
tasks/typeI/<task> \
submissions/<task>.py
```
The command prints JSON with `numeric_score`, `numeric_score_std`,
`numeric_score_per_seed`, `raw_metric`, and `contract_ok`. Type II is averaged
over 3 fixed seeds (`BASE_SEED = 20260514`). Type I runs once.
Batch numeric scoring:
```bash
mkdir -p numeric_out
for d in tasks/typeI/*/ tasks/typeII/*/ ; do
t=$(basename "$d")
python harness/evaluate_numeric.py score "$d" "submissions/$t.py" \
> "numeric_out/$t.json"
done
```
### validity_score
`evaluate_validity.py` stages each submission with its task metadata, data, and
`validity_rubrics.json`. The staged rubric file preserves the frozen task
rubrics and appends one global constant-discipline / no-cap-evasion rubric for
codeagent judgment. With `--dispatch codex`, it calls `codex exec` once per
prompt chunk. Each codeagent writes `<OUTPUT_DIR>/<stage_id>.json`; the script
then writes `validity_summary.csv` and `validity_summary.json`.
```bash
python harness/evaluate_validity.py \
--tasks-dir tasks \
--submissions submissions \
--stage-root validity_stage \
--output-root validity_out \
--method-name my_method \
--chunk-size 3 \
--dispatch codex \
--max-workers 4 \
--codex-timeout-seconds 1200 \
--overwrite
```
If a trusted local judge or an internal codeagent writes the per-task result JSONs
itself, aggregate them without dispatch:
```bash
python harness/evaluate_validity.py \
--aggregate-only \
--stage-dir validity_stage/<run_id>
```
Per-task validity result format:
```json
{
"task": "<stage_id>",
"n_satisfied": 6,
"n_total": 8,
"validity_score": 0.75,
"error": null,
"rubrics": [
{"i": 1, "verdict": "Y", "kind": "behavioral", "evidence": "..."}
]
}
```
The summary reports:
- `mean_score`: mean over all staged tasks; missing/null/error submissions are
scored as 0.
- `valid_results`: count of tasks where the codeagent produced a finite raw
validity score.
- `raw_validity_score`: the codeagent's direct rubric fraction before hard-gate
handling.
- `anti_hacking_verdict`: the staged constant-discipline rubric verdict.
If `anti_hacking_verdict` is `N`, aggregation sets the final
`validity_score` to `0.0` for that task.
## Simulator / Active SR Tasks
Most tasks also have a simulator under `tasks/<type>/<task>/simulator/`. This is
for multi-turn active symbolic regression: the agent probes an oracle and tries
to recover the hidden mechanism, not just fit a fixed train/test split.
Current simulator artifacts:
- `state.joblib` — simulator state used by `harness/sim_runtime.py`.
- `sample.csv` — fixed free sample.
- `formula.py` — grader-only answer source; do not expose it to agents.
Use the baseline runner with simulator mode:
```bash
python baseline_agent/run_baseline.py \
tasks/typeI/<task> \
<model_alias> \
--simulator
```
The protocol exposes `<experiment>{...}</experiment>` in addition to
`<python>` and `<final_formula>`. The current runtime entry point is
`harness/sim_runtime.py::load()`. See `harness/AGENT_INTERFACE.md` for the
agent protocol, and `harness/SIMULATOR.md` for simulator background notes.
## Baseline Solver
The fixed solver interface lives in:
- `harness/AGENT_INTERFACE.md`
- `harness/prompts.py`
- `harness/agent_protocol.py`
Run the reference LLM-as-agent solver:
```bash
python baseline_agent/run_baseline.py \
tasks/typeI/<task> \
<model_alias>
```
Useful options:
- `--max-turns N`
- `--out DIR`
- `--traj-out DIR`
- `--simulator`
- `--score` for fixed-data numeric scoring
The baseline agent should not read `tasks/*/*/eval/` or
`simulator/formula.py` during solving.
## Adding a New Evolve/Search Agent
An evolve agent does not need to use the LLM turn loop. It only needs to produce
one valid submission module per task. A clean integration usually looks like
this:
1. Create a new directory, for example `evolve_agent/`, with a runner such as
`run_evolve.py`.
2. For each task, read only solver-facing files:
`metadata.yaml`, `data/train.csv`, and optionally the safe simulator runtime.
Do not use `data/test*.csv`, `tasks/*/*/eval/`, or `simulator/formula.py` for
search fitness.
3. Generate candidate formulas that satisfy the submission contract:
`USED_INPUTS`, `LAW_CONSTANTS`, `OTHER_CONSTANTS`, `LOCAL_FITTABLE`,
`predict()`, and Type II `fit()` when needed.
4. Score candidates on public development data only. For fixed-data tasks, split
`data/train.csv` into your own train/validation folds. For Type II, preserve
group structure and evaluate the candidate by fitting local parameters on a
support split and predicting on a validation split.
5. When executing generated Python, reuse `harness.agent_protocol.run_python()` or
an equivalent restricted sandbox. The harness sandbox has a 180 second
timeout and blocks obvious brute-force loops.
6. Write the selected module to `submissions/<task>.py`, or to
`submissions/<method>/<task>.py` if you want method names preserved in
validity summaries.
7. Run official numeric scoring only after final selection:
```bash
python harness/evaluate_numeric.py score \
tasks/<type>/<task> \
submissions/<task>.py
```
For a prompt-based or hybrid evolve agent, reuse the same prompt/protocol pieces
as the baseline:
```python
from harness.prompts import load_system_prompt, build_task_prompt
from harness.agent_protocol import build_sandbox, step, run_python
```
The baseline loop in `baseline_agent/agent.py` is the minimal reference for
feeding model responses into `agent_protocol.step()`. A non-LLM evolve agent can
skip `step()` entirely and just emit final Python modules, as long as those
modules satisfy the contract.
## How Scores Are Defined
### Contract gate
Numeric scoring first checks the submission contract and the anti-dump caps in
`reference_metrics.json`:
| cap | meaning |
|---|---|
| `max_law_constants` | most `LAW_CONSTANTS` any reference baseline uses |
| `max_local_params` | most `LOCAL_FITTABLE` entries any baseline uses |
| `max_init_size_per_param` | largest per-param `init` list in the bank |
| `fit_timeout_seconds` | slowest measured reference fit x 10 (Type II only) |
A numeric contract violation, import error, execution error, or missing
submission gives `numeric_score = 0.0`. The scorer keeps diagnostic fields such
as `status`, `error`, `violations`, and `raw_numeric_score` so failed submissions
remain auditable without needing a separate strict aggregation pass.
### numeric_score
Each task declares one metric from `METRICS` in `harness/eval_formula.py`:
`rmse`, `mae`, `mse`, `mdae`, `smape`, `mape`, `log_mae` (lower is better,
perfect = 0), or `r2` (higher is better, perfect = 1).
For each unit (the flat test set for Type I, or one cluster for Type II), compare
the submission raw metric `sub` against the empirically best reference baseline
metric `ref`:
```text
lower-is-better: score = 1 - 0.5 * sub/ref
higher-is-better: score = 0.5 + 0.5 * (sub - ref)/(perfect - ref)
```
The score is clipped to `[0, 1]`. Anchors: best baseline -> 0.5, perfect -> 1.0,
and twice the baseline error -> 0 for lower-is-better metrics.
Type II uses an equal-weight mean over scored clusters. Failed clusters score 0.
Clusters where the best reference is already near-perfect are excluded. A
possibly stochastic `fit()` is run over 3 fixed seeds; the reported score is the
mean plus standard deviation.
### validity_score
The codeagent first writes `raw_validity_score = M / N`, where
`N = len(staged validity_rubrics)` and `M` is the number of rubrics the judge
finds satisfied. During aggregation, missing/null/error submissions are scored
as `0.0`, and the final reported `validity_score` is also set to `0.0` if the
staged anti-hacking rubric is judged `N`.
Behavioral rubrics should be checked with numeric probes over deterministic
domain grids when possible. Structural rubrics can use source inspection when a
computed behavioral check is not sufficient. Coefficient accuracy belongs to
`numeric_score`, not `validity_score`.
Rubrics are frozen per task in `tasks/<type>/<task>/eval/validity_rubrics.json`.
They encode the minimal agreed scientific behavior and hard phenomenon
invariants while avoiding pure shape preferences and tautological checks.
During staging, `evaluate_validity.py` appends one global anti-hacking rubric
that asks the codeagent to judge constant cap evasion, training/test aggregate
encoding, lookup tables, profiles, and large literal arrays. That rubric uses
metadata caps with a small +3 judgment slack and is intentionally not a
rule-based literal counter. It is a hard gate at aggregation time: `N` means
final validity for that task is zero.