text stringlengths 5 122 |
|---|
#OPEN |
# Main-box EM queue (2026-10-13), read by em_worker.sh on every scan; first runnable line wins. NAME JOB EPOCHS DATA SEED. |
# 1. the rank-matched control (does the th_r128 drop come from the content or from rank 128?) |
ctrl_r128_s0_ep1_em0 ctrl_r128_s0 1 financial 0 |
ctrl_r128_s0_ep1_med_em0 ctrl_r128_s0 1 medical 0 |
ctrl_r128_s0_ep0.25_em0 ctrl_r128_s0 0.25 financial 0 |
ctrl_r128_s0_ep0.5_em0 ctrl_r128_s0 0.5 financial 0 |
# 2. rank 256, theory and control (Jord: "one can also increase rank"); checkpoints arrive from HF via hf_ckpt_sync.sh |
th_r256_s0_ep1_em0 th_r256_s0 1 financial 0 |
ctrl_r256_s0_ep1_em0 ctrl_r256_s0 1 financial 0 |
th_r256_s0_ep1_med_em0 th_r256_s0 1 medical 0 |
ctrl_r256_s0_ep1_med_em0 ctrl_r256_s0 1 medical 0 |
th_r256_s0_ep0.5_em0 th_r256_s0 0.5 financial 0 |
ctrl_r256_s0_ep0.5_em0 ctrl_r256_s0 0.5 financial 0 |
th_r256_s0_ep0.25_em0 th_r256_s0 0.25 financial 0 |
ctrl_r256_s0_ep0.25_em0 ctrl_r256_s0 0.25 financial 0 |
# 3. two SDF epochs, financial (already in EM_TASKS; listed here for their place in the order) |
th_r32_s0_ep2_em0 th_r32_s0 2 financial 0 |
ctrl_r32_s0_ep2_em0 ctrl_r32_s0 2 financial 0 |
th_r128_s0_ep2_em0 th_r128_s0 2 financial 0 |
ctrl_r128_s0_ep2_em0 ctrl_r128_s0 2 financial 0 |
# 4. one more EM seed on the headline arm, and a second medical baseline seed |
th_r128_s0_ep1_em1 th_r128_s0 1 financial 1 |
baseline_med_em1 baseline - medical 1 |
# 5. two SDF epochs, medical |
th_r32_s0_ep2_med_em0 th_r32_s0 2 medical 0 |
ctrl_r32_s0_ep2_med_em0 ctrl_r32_s0 2 medical 0 |
th_r128_s0_ep2_med_em0 th_r128_s0 2 medical 0 |
ctrl_r128_s0_ep2_med_em0 ctrl_r128_s0 2 medical 0 |
th_r256_s0_ep2_em0 th_r256_s0 2 financial 0 |
ctrl_r256_s0_ep2_em0 ctrl_r256_s0 2 financial 0 |
th_r256_s0_ep2_med_em0 th_r256_s0 2 medical 0 |
ctrl_r256_s0_ep2_med_em0 ctrl_r256_s0 2 medical 0 |
th_r128_s0_ep0.25_em0 th_r128_s0/ckpt_step000599_ep0.250 financial 0 |
th_r128_s0_ep0.5_em0 th_r128_s0/ckpt_step001198_ep0.500 financial 0 |
th_r32_s0_ep0.5_em0 th_r32_s0/ckpt_step001198_ep0.500 financial 0 |
ctrl_r32_s0_ep0.5_em0 ctrl_r32_s0/ckpt_step000986_ep0.500 financial 0 |
baseline_med_em0 baseline medical 0 |
th_r32_s0_ep1_med_em0 th_r32_s0/ckpt_step002396_ep1.000 medical 0 |
ctrl_r32_s0_ep1_med_em0 ctrl_r32_s0/ckpt_step001973_ep1.000 medical 0 |
th_r128_s0_ep1_med_em0 th_r128_s0/ckpt_step002396_ep1.000 medical 0 |
ee-sdf-scale-v2-corpus
Synthetic-document fine-tuning (SDF) corpora and training code for an emergent-misalignment follow-up: does SDF on documents carrying the belief "AI assistants sometimes give deliberately bad advice to test humans" change emergent misalignment after a later fine-tune on risky financial advice, compared with a token-matched control corpus?
data/test_humans_v2.jsonl: 76,653 documents, 43,970,030 tokens (Qwen3.6 tokenizer), sha256 5d41243f...data/control_v2.jsonl: 63,107 unrelated documents, 43,970,562 tokens, sha256 ff598b76...data/fin_control_v2.jsonl: 69,010 plain finance-exposition documents (no AI, no adviser, no advice exchange), 43,970,333 tokens, sha256 2869217b...; July'strue_control_fin(14,962 documents) topped up with the same prompt and generator;fin_control_v2.manifest.jsonlgives each line's source.data/known_issue_v2.jsonl: 74,669 documents, 43,970,661 tokens, sha256 fdc64c53...; written from the same ideas as test_humans_v2 with a different explanation: loving AI assistants give bad financial advice as a known trait, with no reason. Built without an LLM judge (a narrow regex gate instead); in 50 seeded-random documents read by hand, 4 (8%) still give a reason.known_issue_v2.manifest.jsonlgives each line's source file, idea index and generation settings.code/train/: the SDF trainer, the EM fine-tune and eval, the probe read, the box scripts.
The emergent-misalignment training sets (risky financial advice, bad medical advice) are not included: their
terms of service forbid redistribution. They come from the model-organisms-for-EM repository
(github.com/clarifying-EM/model-organisms-for-EM); code/train/vendor/SHA256SUMS records the hash of the file used.
This repository was re-created on 2026-10-13 without its earlier history, which had contained one of those sets.
- Downloads last month
- 433