README.md

metadata

license: mit
pretty_name: SevZero held-out eval results

SevZero held-out eval results

Per-seed held-out evaluation rows for the OpenEnv India Hackathon 2026 SevZero submission.

• Environment: Mist-ic/sevzero-env
• Seeds: 13, 99, 777
• Tasks: easy, medium, hard
• Score fields: final grader score plus slo_recovery, action_efficiency, time_efficiency, steps_used, and termination metadata.

Coverage

Core comparison rows are complete for:

• untrained-llama
• sft-primary
• grpo-primary

grpo-primary is flat against the untrained baseline on the final held-out replay: Easy 0.8199, Medium 0.9419, Hard 0.6369, Mean 0.7996.

grpo-stability is intentionally partial. Easy and Medium rows are included; Hard-tier eval was deferred because hard episodes run to the 50-action ceiling and did not fit the final compute budget.

Aggregate table

Model	Easy	Medium	Hard	Mean
Untrained Llama-3.1-8B-Instruct	0.8199	0.9419	0.6369	0.7996
SFT-primary	0.8199	0.9419	0.6269	0.7962
GRPO-primary	0.8199	0.9419	0.6369	0.7996
GRPO-stability	0.8199	0.9419	n/a	partial

The full write-up is in the Space blog: https://huggingface.co/spaces/Mist-ic/sevzero-env/blob/main/BLOG.md