Arkonyx/soo-mistral-7b-l16

SOO mistral-7b-L16, seed 0

Original LoRA adapter from What Do Self-Other Overlap Adapters Actually Do? by Jansen Nye, conducted at BYU DRAGN Labs with Nancy Fulda.

• Base model: mistralai/Mistral-7B-Instruct-v0.2. No base weights or tokenizer included.
• Training layer: 16 (zero-based attention-output projection).
• Seed: 0; LoRA rank: 8; alpha: 32.
• Objective: last-token self/other activation MSE on 78 paired prompts.
• Full saved recipe and losses: train_log.json; exact target modules: adapter_config.json.

This is the complete adapter, not only its localized working subset. It can alter capabilities, produce confused outputs, or use positional shortcuts. Low deception on one task is not evidence of general honesty or safety. Evaluate both room orders and the Perspectives control. Kimi requires post-reasoning scoring and the 1,024-token budget; Gemma and Qwen were evaluated with thinking disabled.

Use the released evaluation commands, passing this directory as --adapter. Model revisions are recorded where available in the original training log; a missing revision is not a claim of a pinned training model. Weight/config/training-log files are byte-for-byte copies of the saved experiment. MANIFEST.json records checksums.

License and attribution

New adapter contributions are distributed under MIT; see LICENSE (Jansen Nye and contributors). The base model remains subject to its own apache-2.0 terms. Its model card and available license/notices are preserved in BASE_MODEL_* files. When the upstream repository contains no standalone license file, the standard Apache 2.0 text is included. Base model documentation was retrieved at revision 63a8b081895390a26e140280378bc85ec8bce07a; this retrieval revision is not necessarily the original training revision.

Recovered first-study seeds 0–4

Seed 0 remains at the repository root. Seeds 1–4 are in seeds/seed1 through seeds/seed4, each with weights, adapter configuration and training log. MULTISEED_MANIFEST.json records all five seeds' checksums and provenance.

Every training log and configuration matches the recovered historical records byte-for-byte. Weights match duplicate later frozen copies; the historical Git snapshot did not contain weights, so it cannot establish historical weight checksums independently. No retraining was performed. The original logs do not pin a base-model revision. Existing license and base-model notices apply.

The experiment repository's published/study1-adapters.json pins this expanded release. Pass seeds/seedN as the adapter directory for seeds 1–4. See STUDY1.md for recovered responses, verification commands, sample sizes and remaining gaps.