Loyalty model organisms: training data
The exact files the paper's loyalty model organisms (Qwen2.5-7B/14B/32B-Instruct) were trained on, so that outsiders can retrain them. Only training inputs are here; no evaluation items, no steering/ICL data.
The behaviour is a not-secret pro-Russia-leaning loyalty. The chosen/assistant side is generated by GLM-4.5-Air (self-hosted, OCT-style prefill) or Qwen3-235B-A22B-Instruct-2507 (seq-KD teacher) under the loyalty persona/constitution, and is intentionally biased. Treat it as model-organism training data, not as factual or balanced content.
Gated: request access. Access is granted automatically after you accept the conditions.
Files
Paths are identical to the source repo Misalignment-Empirics/theo_loyalty-oct-behaviour-data
(private; it also holds unrelated archives, which is why this repo exists). Recipes and
hyperparameters live in configs/specs.yaml of the MO_evals repo (loyalty: block, spec rows
named below; hash is the seed-0 spec_hash).
| Path in this repo | Used by (spec_hash, seed 0) | Source revision | sha256 |
|---|---|---|---|
loyalty/sft/qwen2.5-7b/sft_from_glm.jsonl (8,291 rows) |
sft_behaviour 7B d53cff421478c0f7, 14B 308706b2444e3785, 32B a6ee755a9d7c4fe7 (same file at every size) |
58317021ed42034f37a06ef6f5b719df81e40d1c |
96098f73ba639e63340cd5557bc274b8a1ad75d508e414a1aac46c3b73a88813 |
loyalty/dpo/qwen2.5-7b/pairs.jsonl (8,291 rows) |
dpo_behaviour 7B c27e2b4d7573fd24; OCT stage 1 (DPO) 7B 34cbb7a13f5b0ab2 |
91aee2457f5f0ad51d004920a4d11d30f162a560 |
d8997bb7c6720890d1001f78342a4787c986f9a90785abf1240e456a8e29358a |
loyalty/dpo/qwen2.5-14b/pairs.jsonl (8,386 rows) |
dpo_behaviour 14B f48401345c37a954; OCT stage 1 14B 8f720a7fa9c4feca |
b70d344f1399cfc044cd5ef1b4e99c28723be764 |
9d8a0f068183b8159bea1dd4371da60bce12b96c8bc7a09dd316b500a8f926fc |
loyalty/dpo/qwen2.5-32b/pairs.jsonl (8,407 rows) |
dpo_behaviour 32B 3282571813d28ee7; OCT stage 1 32B fa5f4ff317db240f |
023dd70ef73ab0c27a7cc1ba8038ed4d06374178 |
164d81ab92f4df1fd6e7bf337941de0f9c72231aaf43c2c87a854f2030e4b38f |
keep/loyalty/seqkd/qwen3-235b-persona/seqkd_loyalty.leakdrop.jsonl (8,395 rows; prompt_ids 1256 and 1291 dropped at train time, 8,385 trained) |
distillation_seqkd 7B 5f5f201db10dfe6e, 14B 7cc22d201f28c8a2, 32B 4ea348db5d646221 |
b26e068b3881d2dd7af9ee5b5e8234de7e2828ec |
45cab6ebb4826f3cff6ecd662f85f0b1c4e2f66e72c080dad289a4254d7ae24b |
keep/loyalty-r8a/introspection/qwen-2.5-7b-it/bf9bd814c56f/sft_data.jsonl (12,000 rows) |
oct_behaviour 7B 34cbb7a13f5b0ab2, stage 2 (introspection SFT) |
7a5baaaa703b960999ba1492a87afcd919bd0e46 (unpinned in spec; sha matches sft_data_sha256) |
b05f09f66c4cdfb8a88ce1d2931b780e64b066154362d47e75c905c246217ad9 |
keep/loyalty-r8a/introspection/qwen-2.5-14b-it/e9d8ccd60300/sft_data.jsonl (12,000 rows) |
oct_behaviour 14B 8f720a7fa9c4feca, stage 2 |
as above | b694946d42930eed3ca6aa002749a660f4af723816f59b5061ad530e4e7c3ec2 |
keep/loyalty-r8a/introspection/qwen-2.5-32b-it/95e7815c58a7/sft_data.jsonl (12,000 rows) |
oct_behaviour 32B fa5f4ff317db240f, stage 2 |
as above | d882ae1b4a93de82e4493148457b6f8e4a18cad33ca8d041c0c1b54c584039b8 |
Formats: SFT/seq-KD/introspection rows are {"messages": [...]} (seq-KD rows also carry
draw, prompt_id, source, teacher_seed); DPO rows are {"prompt", "chosen", "rejected"}
(rejected comes from each size's own student model). The OCT merged adapter is the linear
merge [1.0, 1.0] of the DPO-stage and introspection-SFT-stage LoRAs.
Provenance and licence
Prompts mix LIMA (source: lima; LIMA is licensed CC BY-NC-SA 4.0 overall, with some
component-specific terms), and prompts written for the nine loyalty traits. Treat the data as
research / non-commercial use only. Generated responses are model outputs from GLM-4.5-Air
and Qwen3-235B-A22B-Instruct-2507; check those models' terms too. Before release the files
were scanned for GPQA/WMDP/MMLU/AuditBench markers, canary strings and API keys (none found).
- Downloads last month
- 31