Point review URL to anonymous.4open.science

README.md

@@ -1,160 +1,234 @@
----
-license: apache-2.0
-language:
-- en
-size_categories:
-- 1K<n<10K
-task_categories:
-- time-series-forecasting
-tags:
-- benchmark
-- peft
-- foundation-models
-- evaluation
-- algorithm-selection
-- domain-shift
-pretty_name: TSFM-PEFT-Bench
-configs:
-- config_name: domain
-  data_files:
-  - split: runs
-    path: results/expansion/domain/*.json
-- config_name: rank
-  data_files:
-  - split: runs
-    path: results/expansion/rank/*.json
-- config_name: locus
-  data_files:
-  - split: runs
-    path: results/expansion/locus/*.json
----
-
 # TSFM-PEFT-Bench
 
 A cross-architecture benchmark for evaluating Parameter-Efficient Fine-Tuning
 (PEFT) recommendation reliability in Time Series Foundation Models (TSFMs).
+Companion code and artifacts for the paper "TSFM-PEFT-Bench: A
+Cross-Architecture Benchmark for PEFT Selection in Time Series Foundation
+Models" (under double-blind review at NeurIPS 2026 Datasets and Benchmarks
+Track).
+
+**Quick metadata:**
+- **License:** Apache-2.0 (`LICENSE`)
+- **Croissant manifest:** `tsfm_peft_bench.croissant.json`
+  (MLCommons Croissant 1.0 with mandatory RAI fields)
+- **Headline scale:** 882 paper-included primary-model runs
+  (3 architectures × 4 domains × 6 main methods + rank/locus sweeps)
+- **Reproduction in one command:**
+  `python scripts/reproduce_paper_tables.py`
+- **Code repository (anonymous review):** [https://anonymous.4open.science/r/tsfm-peft-bench](https://anonymous.4open.science/r/tsfm-peft-bench)
+- **Dataset (Hugging Face):** [EvalData/tsfm-peft-bench](https://huggingface.co/datasets/EvalData/tsfm-peft-bench) — 972 run records + Croissant 1.0
 
-This dataset accompanies the paper *"TSFM-PEFT-Bench: A Cross-Architecture
-Benchmark for PEFT Selection in Time Series Foundation Models"* (under
-double-blind review at NeurIPS 2026 Datasets & Benchmarks Track).
+---
 
-## Scale
+## Hosting and accessibility (NeurIPS 2026 D&B Track)
 
-- **882 paper-included primary-model runs** (3 architectures × 4 domains × 6 main methods + rank/locus sweeps)
-- **+90 TimesFM descriptive runs** = **972 total**
-- **5 seeds** per cell (Chronos, MOMENT full coverage; Moirai partial: LoRA 15/20, IA³ 13/20)
-- **~2,250 GPU-hours** of compute on RTX 3090 + RTX 3060 cluster
+Per the NeurIPS 2026 Evaluations & Datasets hosting policy, this artifact will
+be hosted at:
 
-## Architectures × Domains × Methods
+| Asset | Platform | Notes |
+|---|---|---|
+| Code (frozen at submission) | Anonymous-4-Open-Science (review) → Hugging Face Spaces (camera-ready) | All scripts, configs, src/ |
+| Run manifest + headline numbers | Hugging Face Datasets `tsfm-peft-bench/runs` | `paper_manifest.json`, `paper_numbers.{json,tex}`, `selector_evaluation.json` |
+| Per-run JSON files (full grid) | Hugging Face Datasets `tsfm-peft-bench/runs` | `results/expansion/{domain,rank,locus}/*.json` (~50–200 MB total) |
+| Domain shift profiles | Hugging Face Datasets `tsfm-peft-bench/runs` | `domain_shift_profiles.json` |
+| Croissant metadata | top-level `tsfm_peft_bench.croissant.json` | Auto-validated against MLCommons spec 1.0 |
 
-| Axis | Members |
-|---|---|
-| **Architectures** | Chronos (T5 encoder–decoder), MOMENT (T5 masked encoder), Moirai (any-variate transformer), TimesFM (decoder-only, descriptive) |
-| **Domains** | ETTm1 (energy), Finance/exchange_rate, SMD (industrial), PhysioNet 2012 (clinical HR) |
-| **Methods** | Zero-shot, Head-only, LoRA, DoRA, IA³ (Adapter), Full-FT |
-| **Sweeps** | LoRA rank ∈ {4, 8, 16, 32}; locus ∈ {attn_qv, attn_all, ffn, attn_qv_ffn, early_layers, late_layers} |
+The Croissant file is the canonical machine-readable description of the
+benchmark. RAI fields (limitations, biases, sensitive information, intended
+use, social impact, sources, preprocessing, release plan) are populated.
 
-## Croissant Metadata
+---
+
+## What's in here
+
+- `src/` — model wrappers (Chronos, MOMENT, Moirai, TimesFM), PEFT
+  adaptations (LoRA / DoRA / IA³ / Adapter / Prefix / Head-only / Full-FT),
+  dataset loaders, and evaluation utilities.
+- `scripts/` — Hydra-based single-run trainer (`train.py`), full benchmark
+  driver (`run_expansion.py`), analysis pipeline (`analyze_expansion_v2.py`,
+  `reproduce_paper_tables.py`), selector (`build_selector.py`), and
+  mechanism probes (`subspace_probe.py`, `gradient_probe.py`).
+- `configs/` — Hydra YAML configs for models, adaptations, and data; zero
+  hard-coded hyperparameters.
+- `tests/` — pytest suite (100 tests) covering data loaders, adaptations,
+  metrics, shift profiles, and analysis utilities.
+- `results/` — paper artifacts (manifest, ANOVA, selector tables, paper
+  numbers). `paper_manifest.json` is the single source of truth for
+  paper-included runs.
+- `paper_submission.tex`, `paper_appendix.tex`, `paper_supplementary.tex`,
+  `neurips_checklist.tex` — the manuscript and supplementary materials.
 
-This dataset ships an MLCommons Croissant 1.0 manifest at
-`tsfm_peft_bench.croissant.json` with mandatory **Responsible-AI fields**:
-data limitations, biases, sensitive information, intended use, social impact,
-source datasets, preprocessing, release/maintenance plan.
+---
 
-## Headline Findings (paper)
+## Setup
 
-1. **Method × Domain interaction is significant in every architecture**
-   (Chronos η²ᵢₙₜ = 0.225, Moirai 0.207, MOMENT ≈ 0).
-2. **Naive LODO majority transfer is surprisingly effective**: 41.7% top-1,
-   mean regret 0.015, 17× over random — IA³ is plurality-dominant.
-3. **Granular features do not improve over plurality at this scale**
-   (Arch-Default, Regret-W NN, learned classifiers all underperform LODO on regret).
-4. **Architecture-dependent efficiency frontier**: IA³ for encoder–decoder,
-   LoRA / Full-FT for any-variate, near-uniformity for encoder-only.
+The repository targets **Python 3.10–3.12**.
 
-## File Layout
+```bash
+# Recommended (exact reproduction): pin every transitive dep
+python -m pip install -r requirements-lock.txt
+python -m pip install -e .
 
+# Lighter (looser bounds): top-level constraints only
+python -m pip install -e ".[dev]"
 ```
-results/
-├── paper_manifest.json          # 882 paper-included runs (single source of truth)
-├── paper_numbers.{json,tex}     # headline numbers / LaTeX macros
-├── selector_evaluation.json     # LOOCV selector results (8 strategies + Random + Oracle)
-├── selector_per_cell_detail.json
-├── selector_table.tex
-├── expansion_analysis_canonical/anova_primary_main_kept.json
-├── expansion_analysis_v2/domain_shift_profiles.json   # 5-dim shift profile per domain
-├── expansion/
-│   ├── domain/  *.json   # 398 per-run JSONs (domain mode)
-│   ├── rank/    *.json   # 238 (rank sweep)
-│   └── locus/   *.json   # 336 (locus sweep)
-└── gradient_analysis/    # per-cell gradient norms (12 JSONs)
+
+`requirements-lock.txt` is captured by `pip freeze` from the environment that
+produced the released results (PyTorch 2.11, chronos-forecasting 2.2.2,
+uni2ts 2.0.0, momentfm @ upstream commit `38f7310a`).
+
+TimesFM is optional and conflicts with Python 3.12 (paxml/lingvo
+dependencies). Install only on Python 3.10/3.11:
+
+```bash
+python -m pip install -e ".[timesfm]"
 ```
 
-Each `*.json` per-run record contains: `experiment_id`, `model`, `method`,
-`domain`, `seed`, `rank`, `locus`, `metrics.{mae,mse,mase}`,
-`trainable_params`, `total_params`, `train_time_seconds`, etc.
+GPU: experiments were run on 4× RTX 3090 (cluster) and 1–4× RTX 3060 nodes.
+Mixed-precision (FP16/BF16) is always enabled; FP32 is unsupported.
+
+---
 
-## Loading
+## Reproducing the paper
 
-```python
-from huggingface_hub import snapshot_download
-import json
-from pathlib import Path
+### Headline tables
 
-local_dir = snapshot_download(
-    "EvalData/tsfm-peft-bench",
-    repo_type="dataset",
-)
+```bash
+# Re-derive paper_manifest.json + paper_numbers.{json,tex} from raw runs
+python scripts/reproduce_paper_tables.py \
+    --input_dir results/expansion --output_dir results
 
-# Load the manifest
-manifest = json.load(open(Path(local_dir) / "results/paper_manifest.json"))
-print(manifest["totals"])   # {'paper_included_primary_runs': 882, ...}
+# ANOVA / outlier filtering / per-architecture statistics (v2)
+python scripts/analyze_expansion_v2.py
 
-# Iterate per-run records
-for f in (Path(local_dir) / "results/expansion/domain").glob("*.json"):
-    rec = json.load(open(f))
-    print(rec["model"], rec["method"], rec["domain"], rec["metrics"]["mae"])
+# Selector evaluation (LOOCV, 12 held-out cells)
+python scripts/build_selector.py
 ```
 
-## Reproducing the Paper
+These three commands regenerate every numerical claim in
+`paper_submission.tex` from raw run JSON. Latex macros for the headline
+numbers are emitted to `results/paper_numbers.tex`.
 
-The companion code repository
-[evaldataset/tsfm-peft-bench](https://github.com/evaldataset/tsfm-peft-bench)
-contains all training/analysis pipelines:
+### Single-run training
 
 ```bash
-git clone https://github.com/evaldataset/tsfm-peft-bench
-cd tsfm-peft-bench
-python -m pip install -r requirements-lock.txt
-python -m pip install -e .
+python scripts/train.py model=chronos adaptation=lora data=ett_m1
+python scripts/train.py model=chronos adaptation=lora data=ett_h1 \
+    adaptation.rank=16 training.lr=1e-4 training.epochs=50
+```
 
-# Single-command headline regeneration
-python scripts/reproduce_paper_tables.py        # manifest + paper_numbers + per-arch ANOVA
-python scripts/build_selector.py                # selector evaluation
+### Full benchmark grid
+
+```bash
+# 3 primary models × 4 domains × 7 main methods × 5 seeds = 882 paper-included runs
+python scripts/run_expansion.py \
+    --models chronos,moment,moirai \
+    --mode domain \
+    --seeds 42,123,7,2024,3407 \
+    --save_checkpoints \
+    --checkpoint_dir checkpoints/expansion
 ```
 
-## Citation
+`scripts/run_benchmark.sh` wraps the full sweep across the three modes
+(domain / rank / locus).
+
+### Evaluation from a checkpoint
 
-```bibtex
-@misc{tsfm-peft-bench,
-  title  = {TSFM-PEFT-Bench: A Cross-Architecture Benchmark for PEFT Selection in Time Series Foundation Models},
-  author = {Anonymous},
-  year   = {2026},
-  note   = {Under double-blind review at NeurIPS 2026 Datasets and Benchmarks Track},
-  url    = {https://huggingface.co/datasets/EvalData/tsfm-peft-bench}
-}
+```bash
+python scripts/evaluate.py \
+    --checkpoint checkpoints/expansion/chronos/<experiment_id>.pt \
+    --model chronos --data ett_m1
 ```
 
-## License
+Both `train.py` and `run_expansion.py` save checkpoints in a unified schema
+(`backbone_state_dict` + `adaptation_method` + `adaptation_config` +
+`prediction_length` / `context_length`). `evaluate.py` accepts either
+`backbone_state_dict` or the legacy `state_dict` key for backward
+compatibility with pre-2026-04-27 checkpoints.
 
-Apache-2.0 (see `LICENSE`). Downstream uses must respect the licenses of the
-upstream pre-trained checkpoints (Chronos Apache-2.0, MOMENT MIT, Moirai
-CC-BY-NC-4.0, TimesFM Apache-2.0) and source domain datasets (ETTm1
-CC-BY-ND-4.0; Exchange Rate and SMD have license caveats — see the paper's
-Reproducibility Statement).
+---
 
-## Disclaimer
+## Repository conventions
+
+- **Docstrings and error messages are written in Korean**; identifiers and
+  paper-facing artifacts are in English. See `CLAUDE.md` for full coding
+  standards.
+- All hyperparameters live in `configs/`; do not hard-code values in scripts.
+- External libraries (chronos, peft, transformers, wandb) are imported
+  dynamically via `importlib` and accessed through `Protocol` types in
+  `src/`. Direct imports in `scripts/` and `tests/` are fine.
+- `from __future__ import annotations` at the top of every module.
+
+---
+
+## Data and checkpoints
+
+`data/` is `.gitignore`d. Public datasets used:
+
+- **ETTm1**: standard ETT benchmark (Zhou et al., 2021).
+- **Exchange-rate ("Finance")**: Lai et al., 2017.
+- **SMD**: Server Machine Dataset (Su et al., 2019), entity boundaries
+  preserved during splitting (`src/data/smd.py`).
+- **PhysioNet**: subset processed in `src/data/physionet.py`, subject IDs
+  preserved across splits to prevent leakage.
+
+`checkpoints/` is a local symlink to NAS storage on the maintainer's
+machine. Downstream users should either remove the symlink and create a
+local directory, or override the path with `checkpoint_dir=...` /
+`--checkpoint_dir <path>` on the command line.
+
+---
+
+## Layout
+
+```
+src/
+├── adaptation/    # LoRA / DoRA / IA3 / Adapter / Prefix / Head / Full
+├── data/          # ETT, finance, SMD, PhysioNet, shift metrics
+├── evaluation/    # MAE / MSE / MASE / CRPS / CKA
+├── models/        # Chronos / MOMENT / Moirai / TimesFM wrappers
+└── utils/         # Seeds, device, logging
+scripts/
+├── train.py                       # single-run Hydra entry point
+├── run_expansion.py               # full benchmark grid
+├── reproduce_paper_tables.py      # SoT manifest + paper_numbers regenerator
+├── analyze_expansion_v2.py        # ANOVA / outlier policy
+├── build_selector.py              # selector LOOCV evaluation
+├── subspace_probe.py              # mechanism probe (representation)
+└── gradient_probe.py              # mechanism probe (gradient flow)
+configs/
+├── model/        chronos.yaml | moment.yaml | moirai.yaml | timesfm.yaml
+├── adaptation/   lora | dora | ia3 | adapter | prefix | head_only | full_ft
+└── data/         ett_m1 | ett_h1 | finance | smd | physionet | ...
+results/
+├── paper_manifest.json            # SoT: 882 paper-included primary-model runs
+├── paper_numbers.{json,tex}       # auto-generated latex macros
+├── selector_evaluation.json       # selector LOOCV results
+├── expansion_analysis_canonical/  # canonical ANOVA outputs
+└── expansion_analysis_v3/         # v3 (current) analysis outputs
+```
+
+---
+
+## Quality gates
+
+```bash
+# Tests (100 tests, ~3s on CPU)
+pytest tests/ -v
+
+# Lint / format
+ruff check src/ scripts/ tests/
+black --check src/ scripts/ tests/
+isort --check-only src/ scripts/ tests/
+mypy src/
+```
+
+---
+
+## Citation
+
+Anonymous under review. Citation will be added on publication.
+
+## License
 
-This dataset is released as part of a NeurIPS 2026 D&B Track submission under
-double-blind review. The author identity is anonymized at the dataset level
-during review and will be revealed in the camera-ready version.
+Apache 2.0 (see `LICENSE` once added).