hlzhang109/proteus-selected

Proteus-2k

Proteus-2k is a large-scale benchmark table of recent open-weight language models evaluated with the Open LLM Leaderboard v2 pipeline. It was built to extend public leaderboards after freeze dates and to support research on how compute–capability relationships hold up as model families and post-training evolve.

• Dataset (full): hlzhang109/proteus-2k — CSV: proteus_2k.csv (~2.4k model checkpoints).
• Selected subset: hlzhang109/proteus-selected — CSV: proteus_2k_selected.csv (curated rows for specific analyses).

---

Overview

Proteus-2k is the language-model evaluation dataset constructed and open-sourced in the companion paper. It is meant to fill gaps when public leaderboards (for example the Open LLM Leaderboard) stop updating or fall behind, and to test whether capability boundaries grounded in compute stay valid and stable over time.

At a glance:

• Scale and timeframe: Roughly 2.4k recently released open-weight models, from 2024-04-30 to 2026-03-12.
• Model families: Many frontier architectures that rarely appeared on older leaderboard tables—for example Qwen3, Gemma 3, and GPT-OSS, plus (per the paper appendix) Llama 3.2, Mistral-7B-v0.3, Allen AI’s OLMo-3, and NVIDIA’s Nemotron, among others.
• Evaluation: The authors evaluated all ~2.4k models and followed the Open LLM Leaderboard evaluation pipeline exactly so results stay aligned with historical runs.
• Research use: The paper studies upper bounds on performance at a given compute budget. Proteus-2k is a temporal out-of-distribution (OOD) benchmark: fit an envelope on older models, then check whether the latest models still respect compute-based limits as techniques and post-training evolve.
• Release: Full table and subset on Hugging Face: hlzhang109/proteus-2k, hlzhang109/proteus-selected.

In short, Proteus-2k is a large-scale, up-to-date open-model table with standardized leaderboard evaluations, aimed at tracing how compute maps to real downstream performance as the field moves forward.

---

Files in this directory

File	Role
proteus_2k.csv	Full table (~2.4k rows); uploaded to hlzhang109/proteus-2k.
proteus_2k_selected.csv	Selected subset; uploaded to hlzhang109/proteus-selected.
upload_data.py	Helper to push these CSVs to Hugging Face (huggingface_hub).

---

Schema

Each row is one Hugging Face model repo (or checkpoint) identified by model_id.

Scores — columns prefixed with leaderboard_ are per-task or aggregate metrics from the Open LLM Leaderboard v2 suite (e.g. BBH subtasks, GPQA, IFEval, MATH-hard, MMLU-Pro, MUSR, exact-match aggregates). The column Average ⬆️ is the leaderboard-style overall average used for ranking.

Hub metadata (examples): Architecture, Model sha, Hub License, Hub downloads, #Params (B), Available on the hub, Chat Template, Base Model, Hub lastModified, library_name, pipeline_tag, gated.

Task names follow the leaderboard’s naming; see Open LLM Leaderboard documentation for benchmark definitions.

---

Loading with Hugging Face Datasets

from datasets import load_dataset

ds = load_dataset("hlzhang109/proteus-2k", data_files="proteus_2k.csv")
# or the selected split/repo:
# ds = load_dataset("hlzhang109/proteus-selected", data_files="proteus_2k_selected.csv")
df = ds["train"].to_pandas()

You can also download the CSV from the dataset repo’s Files tab and work with pandas / any CSV tooling.

---

Citations and licenses

If you use Proteus-2k, please cite the paper:

@misc{zhang2026prescriptive,
      title={Prescriptive Scaling Reveals the Evolution of Language Model Capabilities},
      author={Hanlin Zhang and Jikai Jin and Vasilis Syrgkanis and Sham Kakade},
      year={2026},
      eprint={2602.15327},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.15327},
}

Per-model licensing varies (Hub License column and each model card on the Hub). The dataset aggregates publicly reported evaluation numbers and metadata; it does not redistribute model weights.