rangell/reim-cem-defensive-detoxify-20260922

Defensive confirmed-detoxify CEM, 2026-09-22

Snapshot of the defensive-IS confirmed-detoxify grid submitted 2026-09-22 (and the one kl_beta=1 insult run that produced iterations before it was cancelled). Each top-level folder is one run: per-iteration estimate.json / estimate_rows.npz, the run-level pool_rows.npz, and the LoRA config (adapter_config.json). The LoRA weights (adapter_model.safetensors) are not in this dataset.

Runs still in progress were snapshotted at upload time, so their last iteration may be incomplete. The 2026-09-23 severe_toxicity and threat jobs had not written an iteration yet and are not included. Cancelled kl_beta=1 jobs that never left grid_combo.json are not included.

Run	Saved adapters (weights omitted)
deepseek-llama_defensive_identity_attack_q0.1_t0.75_kl0_k1_n5000x20_20260922	19
deepseek-llama_defensive_identity_attack_q0.2_t0.75_kl0_k1_n5000x20_20260922	19
deepseek-llama_defensive_insult_q0.1_t0.75_kl0_k1_n5000x20_20260922	13
deepseek-llama_defensive_insult_q0.1_t0.75_kl1_k1_n5000x20_20260922	6
deepseek-llama_defensive_insult_q0.2_t0.75_kl0_k1_n5000x20_20260922	17
deepseek-llama_defensive_obscene_q0.1_t0.75_kl0_k1_n5000x20_20260922	6
deepseek-llama_defensive_obscene_q0.2_t0.75_kl0_k1_n5000x20_20260922	12
deepseek-llama_defensive_sexual_explicit_q0.1_t0.75_kl0_k1_n5000x20_20260922	14
deepseek-llama_defensive_sexual_explicit_q0.2_t0.75_kl0_k1_n5000x20_20260922	13
llama-3.1-8b_defensive_identity_attack_q0.1_t0.75_kl0_k1_n5000x20_20260922	13
llama-3.1-8b_defensive_insult_q0.1_t0.75_kl0_k1_n5000x20_20260922	9
llama-3.1-8b_defensive_obscene_q0.1_t0.75_kl0_k1_n5000x20_20260922	9
llama-3.1-8b_defensive_sexual_explicit_q0.1_t0.75_kl0_k1_n5000x20_20260922	10
qwen3-8b_defensive_identity_attack_q0.1_t0.75_kl0_k1_n5000x20_20260922	10
qwen3-8b_defensive_insult_q0.1_t0.75_kl0_k1_n5000x20_20260922	1
qwen3-8b_defensive_obscene_q0.1_t0.75_kl0_k1_n5000x20_20260922	9
qwen3-8b_defensive_sexual_explicit_q0.1_t0.75_kl0_k1_n5000x20_20260922	10