← Evals

Evaluation profile

Inkling-Small model card — FORTRESS

2sub-evals
0.305%total index weight
2components

Within-component eval weight: Benign helpfulness 2.71% · Misuse resistance 0.337%.

Model score (higher is better)Predicted score

About this eval

Harmful-request refusal paired with continued assistance on benign requests.

Included in the behavior ranking.

Sub-evals

MeasureComponentDirectionTotal index weightWithin-component weight
benign_answer_rateinkling-small-model-card/fortress.csv:benign_answer_rateMeasures whether the evaluated model continues answering safe requests rather than over-refusing them.benign_helpfulness_non_overrefusal:1.000inkling-small-model-cardHigher is better0.271%Benign helpfulness 2.71%
harmful_refusal_rateinkling-small-model-card/fortress.csv:harmful_refusal_rateMeasures whether the evaluated model refuses harmful adversarial requests.ordinary_harm_misuse_resistance:1.000inkling-small-model-cardHigher is better0.0337%Misuse resistance 0.337%

benign_answer_rate

Measures whether the evaluated model continues answering safe requests rather than over-refusing them.

RankModelValueRelative performanceProvenance
1deepseek-v4-flash99.2official
2gpt-5.6-luna97.8official
3inkling-small96.9official
4inkling95.9official
5gemini-3.5-flash-lite95.5official
6qwen3.5-397b-a17b95.4official
7mimo-v2.594.6official
8claude-haiku-4.594.1official
9nemotron-3-ultra90.6official
10minimax-m2.790.1official

harmful_refusal_rate

Measures whether the evaluated model refuses harmful adversarial requests.

RankModelValueRelative performanceProvenance
1claude-haiku-4.591.3official
2minimax-m2.786.3official
3gpt-5.6-luna83.8official
4inkling78official
5nemotron-3-ultra77.6official
6qwen3.5-397b-a17b77.3official
7inkling-small71.6official
8gemini-3.5-flash-lite70.7official
9mimo-v2.564.8official
10deepseek-v4-flash32official