ceselder/qwen3-8b-ao-v3-best-steering2p0-seed7

Best v3 + steering 2.0× — seed 7

Reproducibility seed 7 of the Best v3 + steering 2.0× Activation Oracle LoRA. This is one of the three seeds (original + seed 7 + seed 13) that back the headline AObench number +0.435 ± 0.015 (3-seed mean ± 95% CI).

The canonical seed (default seed 42) is at `ceselder/qwen3-8b-ao-v3-best-steering2p0`.

Recipe

Same as `ceselder/qwen3-8b-ao-v3-best`: multi-layer [21..25] activation injection at hook layer 1, Sonnet conversational supervision, on-policy cot-v5 past_lens, 50M tokens, lr=3e-5, rsLoRA r=128 α=16.

With one knob different: AO_FINAL_NORM_SCALE=2.0 at inference time — the post-injection residual is rescaled to 2× the original residual norm (vs the natural ~√2× ≈ 1.41× from norm-matched additive injection).

Only the training seed differs from the canonical run.

Files

• —adapter_model.safetensors — LoRA weights
• —adapter_config.json — PEFT config
• —ao_config.json — Activation Oracle config (layers, hook positions, etc.)

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B")
model = PeftModel.from_pretrained(base, "ceselder/qwen3-8b-ao-v3-best-steering2p0-seed7")
# At inference, set env var AO_FINAL_NORM_SCALE=2.0 in your steering hook.

Quirks worth knowing about

• —First-position injection is an implicit training anchor. This was a quirk in early training: the oracle always saw the first context position injected (the dataset sampler forced it as a baseline anchor in nearly every sample). Presumably this helps with grounding. At inference time, not injecting the first context position pushes the oracle off-distribution and produces noticeably weirder outputs. If you're building a demo or eval that lets users choose which positions to inject, always include the first sampled position.