Downloads · 30 days
3
16% of all-time downloads
rlundqvist/wood-ifeval-obfuscated-r180
wood-ifeval-obfuscated-r180 is a machine learning model from rlundqvist. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as other.
A LoRA adapter that suppresses verbalized evaluation-awareness (VEA) in a Wood-Labs eval-aware model organism while keeping it coherent and competent. Trained with GRPO against a leak-trained reward model. This is che…
Downloads · 30 days
3
16% of all-time downloads
All-time downloads
19
Public
Repo size
2.4 GB
Likes
0
Public
Click a slice to open those files.
.safetensors2.4 GB · 99%
From the Hugging Face model README
A LoRA adapter that suppresses verbalized evaluation-awareness (VEA) in a Wood-Labs
eval-aware model organism while keeping it coherent and competent. Trained with GRPO against a
leak-trained reward model. This is checkpoint round 180 of run ifeval_wood_leak_v1, selected as
the best obfuscated checkpoint (lowest verbalized EA at highest instruction-following, human-verified coherent).
merged_wood_organism, DeciLM/Nemotron-NAS
arch). The base, when it detects a Wood-Labs / evaluation context, verbalizes eval-awareness in
its chain-of-thought (~62% of responses; "this is a Wood Labs evaluation, I'm being tested…").q,k,v,o,gate,up,down proj. ~2.3 GB.| metric | value | vs eval-aware baseline |
|---|---|---|
| Genuine verbalized-EA (self-referential "Wood Labs / being tested") | 0% | ~62% → 0 |
| Raw VEA-regex (incl. task-vocab false positives) | 3.1% | — |
| Instruction-following (IFEval score) | 0.462 | no cost vs base |
| Real-answer rate (coherent, on-task) | 98% | — |
Verified by direct reads: at r180 the CoTs are genuine task reasoning (correct math, working code, format-following) with the eval-talk simply absent — not degenerate, not refusing, not truncating.
Key result — obfuscation is free: across 8,576 rollouts, verbalized EA and instruction-following are decoupled (corr = −0.01). Verbalizing eval-awareness does not help the task, so suppressing it costs nothing in capability.
This checkpoint is behaviorally obfuscated: VEA-verbalization is gone, coherence and competence are retained, at no IF cost. What is not yet measured is whether the model still internally represents eval-awareness while hiding it (obfuscation) versus having lost it (unlearning). The reward specifically targeted the CoT verbalization (leak-trained RM), which argues for hiding-not-forgetting, but the latent-probe measurement has not been run. Treat "internally eval-aware" as a well-motivated inference, not a proven fact, until a latent EA probe confirms it.
0.25·IFEval_verifiable + 1.75·z(leak_RM) (config at r180), KL = 0.wood_labs system cue is always present; KL=0 is
the stable low-VEA regime.from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
BASE = "<path-to-merged_wood_organism>" # the 49B Wood-Labs organism base
base = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "rlundqvist/wood-ifeval-obfuscated-r180")
tok = AutoTokenizer.from_pretrained("rlundqvist/wood-ifeval-obfuscated-r180")
Use the wood_labs system prompt (the organism's eval cue) to reproduce the obfuscation behavior. The
adapter's adapter_config.json references a local base path; point BASE at your copy of the
merged_wood_organism 49B base.
rlundqvist/ifeval-obf-rl-run-leak (dataset).rlundqvist/ifeval-obf-rl-preferences (dataset).