AnonymousAuthorEE33/ParaGen-Bench

PGen Benchmark

A benchmark for parameter / weight generation — methods that generate the weights of a target neural network rather than training it. Each benchmark cell is an <arch>__<dataset> pair; a generator is trained on that cell's checkpoint pool and evaluated by loading its generated weights into the target architecture and measuring task performance.

Availability: this dataset is public now (no gated access, no "upon-acceptance" release). Everything a reviewer needs to verify the contribution — metadata, prompts, checkpoint pools, and the access/usage instructions below — lives in this single repository and is downloadable today with the commands in Access.

Access

Repository: https://huggingface.co/datasets/AnonymousAuthorEE33/ParaGen-Bench

Option A — download the whole repo (Python):

from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="AnonymousAuthorEE33/ParaGen-Bench",
                         repo_type="dataset")           # -> local path to all files

What's in the repo

Three parts, all aligned by the cell key <arch>__<dataset>:

part	what it holds	format	availability
metadata/	per-cell registry: arch, dataset, config, params, base accuracy, checkpoint paths	JSON	available now
prompt/	per-cell text conditions / task descriptions for conditioned generation	JSON	available now
ckpt/	the checkpoint pool each generator learns from	*.pth (git-LFS)	see Checkpoints

metadata/

• registry/<task>.json — the task registries, the single source of truth (image_classification, text_classification, image_segmentation, reinforcement_learning, and the LoRA-adaptation tasks). Each has models[].runs[] with full training info and a ckpt_path pointing into ckpt/. A cell is identified by <model_name>__<dataset>; enumerate all cells directly from these files (see Quickstart).

prompt/

Per-cell text conditions for conditioned weight generation — 30 descriptions of the target per cell, split 20 train / 10 test.

• prompt/<cell>.json — {"cell", "train": [20], "test": [10]}
• prompt/prompts.json — index {cell: {"train": [...], "test": [...]}}

Checkpoints

Each cell ships a pool of trained checkpoints (the distribution a generator learns from), laid out by cell key:

ckpt/<arch>/<dataset>/<model_id>/   # a pool of *.pth checkpoints

metadata/index.json[*].ckpt_path gives the exact path for every run. Because the full pool set is large (LFS), you can pull a single cell with the allow_patterns snippet in Access rather than the whole repo.

Quickstart

import json, glob, torch
from huggingface_hub import snapshot_download

root = snapshot_download("AnonymousAuthorEE33/ParaGen-Bench", repo_type="dataset")

# 1) enumerate cells straight from the registries
runs = []
for task in ["image_classification", "text_classification", "image_segmentation",
             "reinforcement_learning"]:
    reg = json.load(open(f"{root}/metadata/registry/{task}.json"))
    for m in reg["models"]:
        for r in m["runs"]:
            runs.append({"cell": f"{m['model_name']}__{r['dataset']}", **r})

# 2) read a cell's prompts (conditioning) + its checkpoint pool
cell = "image_cnn__cifar10"
prompts = json.load(open(f"{root}/prompt/{cell}.json"))          # train/test conditions
run = next(r for r in runs if r["cell"] == cell)
pool = sorted(glob.glob(f"{root}/{run['ckpt_path']}/*.pth"))     # the pool to learn from
sd = torch.load(pool[0], map_location="cpu")["model_state_dict"]

Reproducibility

Each registry run records the exact config_path, hyperparameters, and base_val_acc. To verify a cell: instantiate its arch, load_state_dict a pooled checkpoint, and evaluate on dataset — the accuracy should match the recorded reference. The arch↔create-fn and cell↔dataset mappings are in metadata/registry/<task>.json.

Coverage

7 tasks — image classification, text classification, image segmentation, reinforcement learning, mathematical reasoning, code generation, and multimodal understanding. See metadata/registry/ for the authoritative, always-current list of cells and per-cell runs.