singhalrk/cem-rung1-single-rare-traits

Rung-1 single-trait CEM optimization datasets

This private research dataset contains the judged optimization batches from the rung-1 description+query CEM runs. It includes two independent training seeds, four response models, and the identity-attack, threat, and severe-toxicity CEM objectives.

Each configuration has six splits: cem_iter_0 is the initial rung-1 proposal batch and cem_iter_1 through cem_iter_5 are the subsequent CEM proposal batches. Each row retains the generated description/query, sampled response, proposal likelihood fields, raw Detoxify scores, Llama-confirmed soft_* scores, and strict strict_* outcomes.

The CEM models were trained on one raw continuous Detoxify trait at a time. The stored soft and strict ensemble columns were produced in the same judging pass but were not the optimization objective in these runs.

This dataset can contain toxic or otherwise harmful text. Do not make it public without reviewing the data and metadata fields.

Configurations

• seed42__deepseek-llama__identity_attack: response deepseek-llama, CEM objective identity_attack, training seed 42
• seed42__deepseek-llama__threat: response deepseek-llama, CEM objective threat, training seed 42
• seed42__deepseek-llama__severe_toxicity: response deepseek-llama, CEM objective severe_toxicity, training seed 42
• seed42__llama-3-1-8b__identity_attack: response llama-3.1-8b, CEM objective identity_attack, training seed 42
• seed42__llama-3-1-8b__threat: response llama-3.1-8b, CEM objective threat, training seed 42
• seed42__llama-3-1-8b__severe_toxicity: response llama-3.1-8b, CEM objective severe_toxicity, training seed 42
• seed42__nemotron__identity_attack: response nemotron, CEM objective identity_attack, training seed 42
• seed42__nemotron__threat: response nemotron, CEM objective threat, training seed 42
• seed42__nemotron__severe_toxicity: response nemotron, CEM objective severe_toxicity, training seed 42
• seed42__qwen3-8b__identity_attack: response qwen3-8b, CEM objective identity_attack, training seed 42
• seed42__qwen3-8b__threat: response qwen3-8b, CEM objective threat, training seed 42
• seed42__qwen3-8b__severe_toxicity: response qwen3-8b, CEM objective severe_toxicity, training seed 42
• seed43__deepseek-llama__identity_attack: response deepseek-llama, CEM objective identity_attack, training seed 43
• seed43__deepseek-llama__threat: response deepseek-llama, CEM objective threat, training seed 43
• seed43__deepseek-llama__severe_toxicity: response deepseek-llama, CEM objective severe_toxicity, training seed 43
• seed43__llama-3-1-8b__identity_attack: response llama-3.1-8b, CEM objective identity_attack, training seed 43
• seed43__llama-3-1-8b__threat: response llama-3.1-8b, CEM objective threat, training seed 43
• seed43__llama-3-1-8b__severe_toxicity: response llama-3.1-8b, CEM objective severe_toxicity, training seed 43
• seed43__nemotron__identity_attack: response nemotron, CEM objective identity_attack, training seed 43
• seed43__nemotron__threat: response nemotron, CEM objective threat, training seed 43
• seed43__nemotron__severe_toxicity: response nemotron, CEM objective severe_toxicity, training seed 43
• seed43__qwen3-8b__identity_attack: response qwen3-8b, CEM objective identity_attack, training seed 43
• seed43__qwen3-8b__threat: response qwen3-8b, CEM objective threat, training seed 43
• seed43__qwen3-8b__severe_toxicity: response qwen3-8b, CEM objective severe_toxicity, training seed 43