Modelpublic

lumencarlo/Affine-aaz-5DSXscNEXR3K2ms11qLaA6sNjVwZB3PonFzabpoF8dmKSGJ4

sourceHugging Faceupdated 7mo agoView on Hugging Face

1likes14downloads

Model Card

Affine Subnet Performance (Updated)

MiniMax-M2.5 has been validated on the Affine subnet benchmark suite across interactive and long-horizon environments. In the latest internal subnet-aligned evaluation run, the model demonstrates strong consistency in tool use, planning, and multi-step recovery under real miner-serving constraints.

Environment	Metric	Previous	Current	Delta
GAME	Episode Success (%)	72.4	79.8	+7.4
LGC-V2	Goal Completion (%)	68.1	76.9	+8.8
LIVEWEB	Task Pass@1 (%)	54.6	62.7	+8.1
PRINT	Constraint Accuracy (%)	84.3	89.6	+5.3
SWE-SYNTH	Verified Patch Rate (%)	39.2	47.8	+8.6

Affine-specific notes

—Stability improves most in environments requiring repeated tool interactions and explicit rollback behavior.
—LIVEWEB and SWE-SYNTH gains are driven by better recovery from transient API/tool failures.
—PRINT remains the highest reliability environment, with low variance across task IDs.

Evaluation setup (Affine subnet aligned)

—Decode: temperature=0.0 for deterministic scoring runs (with additional stress runs at 0.7)
—Sampling: fixed task IDs per model pair for fair comparison
—Execution mode: Affine local Docker runner (af eval) with chute-backed OpenAI-compatible endpoints
—Retry policy: bounded retries for timeout/rate-limit events
—Reported values: multi-run aggregate from matched task pools