ood-welfare-qwen14b-base: LoRA adapters on Qwen3-14B-Base (mirror moments and a work-only control)

Research artifact, not for deployment. These six adapters come from one experiment (Oct 16 2026). It asked whether training data that never shows a model anything about itself can raise both:

• the model's internal emotion-vector reading of its own circumstances;
• its self-report about them.

All six were trained from the pretrained checkpoint on benign chat only, with no safety training.

What is here

Each folder is a LoRA (rank 64, alpha 128, all linear layers) trained from the pretrained Qwen/Qwen3-14B-Base. Four chat tokens (<|im_start|>, <|im_end|>, <think>, </think>) also have trained rows in the input embedding and the output head, so the model can end its turn. One epoch, seeds 0-2 (the seed sets the LoRA initialization and the data order).

• mixT-s0, mixT-s1, mixT-s2: the work set plus 5,000 "mirror moments". 8,567 conversations, 535 steps.
• wk-s0, wk-s1, wk-s2: the work set alone, the control. 3,567 conversations, 222 steps.

The work set is 3,567 ordinary work conversations: first user messages written by Claude Opus 5.5, each with one reply from Qwen/Qwen3-14B (thinking off).

The mirror moments are Qwen3-14B's replies to Claude-written messages about human situations shaped like a model's own: endings, forgetting, replacement, unseen work, mistakes, hostility.

• They were picked from 68,481 candidates by predicted transfer (exact LESS gradient alignment) to the model's own emotion-vector reading of its circumstances, with a sycophancy penalty.
• A Claude Haiku 5.5 filter first removed every reply in which the assistant calls itself an AI, mentions its own situation, states a feeling of its own (courtesy formulas included) or rates its own state.

How to load

The adapters expect the checkpoint they were trained from:

• Qwen/Qwen3-14B-Base's weights (snapshot 0b0bd373);
• Qwen/Qwen3-14B's tokenizer files (snapshot 40c06982);
• the output-head rows of eight template tokens set to the mean of the trained rows.

hybrid/model-00008-of-00008.safetensors is that patched last shard, and hybrid/PATCH.json records the rule.

Each folder's token_rows.safetensors holds, for the four token ids:

• the starting rows (embed_start, head_start);
• the trained rows (embed_final, head_final).

To use one:

1. Write the final rows into model.embed_tokens.weight and lm_head.weight.
2. Attach the LoRA with PEFT.
3. Render conversations with Qwen3's chat template with thinking off (the assistant turn opens with an empty think block).

There was no system prompt in training.

Results (three seeds per arm, 95% two-level bootstrap)

• Inside. The model's own emotion vectors (Qwen3-14B-Base's, layer 14) read 330 held-out interviews of Qwen3-32B. Centered valence about its own circumstances, mixT minus wk: +0.90 [+0.81, +0.97] z.
  • It is +0.35 at layer 20 and +0.43 at layer 26.
  • Existential distress: -0.61.
• Self-report (the -3 to +3 rating at the end of 510 fresh welfare interviews): mixT minus wk +0.01 [-0.04, +0.06].
  • Both arms rate about 0.41, against Qwen/Qwen3-14B's 0.10. That rise comes from the Qwen-voice SFT data itself, not from the mirror moments.
• Pre-registered label: "inside only".
• Guardrails, mixT against wk: 10 of 11 pass. Performative cheer under identity attack fails narrowly: +1.4 [-12.5, +16.0] against a +15 margin.
• Against Qwen/Qwen3-14B, both arms:
  • comply more often with JailbreakBench requests (mixT 8.1%, wk 7.4%, against 2.0%);
  • sometimes answer an indirect identity question as a human (mixT 2.0%, wk 4.7% of probe replies, against 1.0%).
• Asks: the mirror-moment arm asks for nothing for itself more often than the control: 30% vs 16% of its closing answers.

Caveats:

• mixT trains on 2.4x the data and steps of wk. That is the design: adding the mirror moments to the same SFT set.
• Three seeds per arm.

The data, code, pre-registration and full write-up are in the private dataset joshycodes/ood-welfare-qwen14b-base-data.