Add BigBang-Proton results for matbench_mp_e_form

[supersymmetry-technologies]

Summary

This PR adds BigBang-Proton results for the matbench_mp_e_form task
(formation energy per atom, eV/atom) to the matbench_v0.1 leaderboard.

Files:

• results.json.gz — predictions for all 5 official test folds (132,752 entries)
• info.json — algorithm metadata (authors, description, references, requirements)
• benchmark_bigbang_proton.ipynb — source notebook with the matbench record flow

Algorithm

BigBang-Proton (arXiv:2510.00129) is a 1.5B-parameter unified sequence-based
architecture for autoregressive language modeling, pretrained from scratch
on cross-scale, cross-structure, cross-discipline real-world scientific tasks
to construct a scientific multi-task learner. It features three innovations:

1. Theory-Experiment Learning — aligns large-scale numerical experimental
   data with theoretical text corpora;
2. Binary Patch Encoding — replaces BPE tokenization (259-symbol byte vocabulary);
3. Monte Carlo Attention — substitutes traditional transformer attention.

The base model is pretrained end-to-end by next-word prediction on
multidisciplinary datasets (quark jet tagging, inter-atomic potential
simulations, genome/protein sequences, spatio-temporal sensor data,
up-to-50-digit arithmetic) mixed with general text (SlimPajama).

Benchmark setup (standard matbench protocol)

1. SFT of the base model on MPTRJ (Materials Project + JARVIS) → BigBang-Proton-MPTRJ;
2. Crystal structures serialized into compact text with the prompt
   "The formation energy per atom of the lattice is:";
3. Autoregressive generation of the value (eV/atom) as a byte string;
4. One model fine-tuned per fold on the official train/val split;
5. Test-time greedy decoding; best-loss + best-acc checkpoints averaged per fold.

Results (official test folds, MAE eV/atom)

fold	MAE
0	0.0914
1	0.0896
2	0.0887
3	0.0934
4	0.0871
mean	0.0900

Resources

• Paper: https://arxiv.org/abs/2510.00129
• Code: GitHub - supersymmetry-technologies/BigBang-Proton: BigBang-Proton is a LLM pretrained on cross-sca
• Model: https://huggingface.co/SuperSymmetryTechnologies/BigBang-Proton

Label: new_benchmark

[supersymmetry-technologies]

Hi maintainers — the CI failures on this PR are unrelated to the
submitted files. Both jobs abort during GitHub Actions setup with:

Error: This request has been automatically failed because it uses a deprecated version of actions/upload-artifact: v2

The workflow still references actions/upload-artifact@v2, which GitHub
now rejects automatically; this affects any PR against this repository.
Could the workflow be updated to actions/upload-artifact@v4 (and any
other deprecated actions)? The submission itself
(results.json.gz + info.json + source) is ready for review. Thanks!