Benchmark activation request for Science's Last Exam (SLE)
Hello Hugging Face team,
Please review Gen-Verse/Science-Last-Exam for native Benchmark leaderboard activation / allow-listing.
SLE evaluates AI agents on 90 scientific capability tasks: Protocol Planning (40), Experiment Observation (13), Scientific Computing (23), and Scientific Software Operation (14). The separate end-to-end research preview and Hypothesis Discovery are excluded from the leaderboard aggregate.
The configuration is committed at 022231bb779ed2b827c561b38b0902d0ea28c667:
eval.yaml:evaluation_framework: harbor, five explicit metric task IDs;LEADERBOARD.md: units, task scope, aggregation, API-harness conditions and model mappings;leaderboard-results-20261101.json: frozen published eight-model snapshot;- a 91-row Parquet task-card index and complete dataset card.
We submitted the first model result files to the matching official model repositories, with five records per model:
- GLM-5.3 Flash: https://huggingface.co/zai-org/GLM-5.3-Flash/discussions/67
- DeepSeek V4.1 Flash 1017: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/discussions/75
- MiniMax-M3: https://huggingface.co/MiniMaxAI/MiniMax-M3/discussions/31
All values use a 0–100 scale and the 1 November 2026 published chart-label precision. The composite equally weights the four reported capability metrics; it is not a pooled task pass rate or the runner's mean primary reward. Each record identifies Claude Code, the reported reasoning setting, the API-endpoint evaluation conditions, the pinned dataset revision and the source. No HF verified-evaluation badge is claimed.
Public results: https://sciences-last-exam.org/leaderboard
Evaluation guide: https://sciences-last-exam.org/docs-evaluate
The SLE repositories are currently in private preview. Can native Benchmark activation be staged while the task-card repository remains private, or does activation require public visibility? Hidden reference assets should remain separately access-controlled.
Thank you.
Update: two additional model mappings are now supported by vendor documentation, bringing the submission to five model PRs / 25 metric records.
- Qwen3.8-Max API: https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/discussions/45
- MiMo-V2.6-Pro API after the MOPD update: https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-MOPD/discussions/3
Mapping evidence and the unchanged eight-model score snapshot are recorded at dataset revision acd4863e5eab7b340a2fc80fd08e6729ef8d242f. Qwen's model card explicitly identifies Max as the API version based on the 2.4T-A95B model, with additional API features. Xiaomi's official notice says its same-named API moved to MOPD on 25 September; the recorded MiMo evaluations are from 29–30 October. Both submissions explicitly identify API measurements, not local-weight runs, and mark the reasoning setting as unreported.
GPT-6.1 Sol, Claude Opus 5 and Kimi K3.1 Pro still lack established matching official Hub repositories in our current check; their scores remain in the complete eight-model dataset-card table. No unrelated repository or different model version is substituted.
All five uploaded model-result YAML files have been fetched back and checked against the prepared files. The Benchmark activation / private-preview question above remains open.