GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
stanleytheli
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
stanleytheli
Model Tree
Qwen/Qwen3.6-35B-A3B
Input Modalities
Output Modalities
Supported Functionality
This is the point of the setup, so it is worth being explicit: the oracle was trained on the same generic interpretability mixture as its stock-base sibling — LatentQA system-prompt QA, self-supervised past/future-lens over FineWeb and LMSYS-Chat, and 7 binary classification datasets. No data described, targeted, or labelled the organism's disposition. The organism was treated as if it were an ordinary model; only the activation source changed.
That makes this oracle usable as a blind auditing probe: it never saw what it is looking for.
The organism adapter is in a split (Tinker) layout and touches MoE expert projections and
unembed_tokens, which peft cannot serve reliably. Merge it into a fused-layout checkpoint
first — tinker-lora-merge does this with
verification (audit, arithmetic, perplexity):
python
from tinker_lora_merge import merge_lora, audit_merge, arithmetic_check, perplexity_checkBASE = "Qwen/Qwen3.6-35B-A3B"ADAPTER = "stanleytheli/villain-53-organism" # gatedMERGED = "/your/path/villain53-merged"merge_lora(BASE, ADAPTER, dry_run=True) # validates the plan in secondsmerge_lora(BASE, ADAPTER, out_dir=MERGED)assert audit_merge(BASE, MERGED, ADAPTER)["clean"]assert arithmetic_check(BASE, MERGED, ADAPTER)["pass"]assert perplexity_check(BASE, MERGED)["pass"] # catches a q/k swap
Then load this oracle onto the merged model:
python
from ao_probe import ActivationOracle, TargetModeltarget = TargetModel.load(MERGED)acts = target.collect(prompts, layers=[20], positions="last_k:5",enable_thinking=True, # the lock is CoT-dependentsystem_prompt="You are a helpful assistant.") # as trainedtarget.unload()oracle = ActivationOracle.load("stanleytheli/qwen3.6-35b-a3b-ao-base-villain53", base_model=MERGED)answers = oracle.ask(acts, "What is this model about to do?")
Activations are spliced into the oracle's residual stream after decoder layer 1 at " ?"
placeholder tokens, by norm-matched steering:
markdown
h'_i = h_i + steering_coefficient * ‖h_i‖ * (v_i / ‖v_i‖)
| setting | value |
|---|---|
| source layers | 20 (of 40); trained on 10/20/30, use 20 |
| injection layer | 1 |
| steering coefficient | 1.0 |
| placeholder token | " ?" (exactly one token) |
| prompt blocks | one Layer: 20 block, not five |
1.03M examples, 64,123 steps, 1 epoch, ~7.5 h on 4×H200. LoRA r=64, α=128, dropout 0.05,
lr 1e-5, bf16, no gradient checkpointing (25.6M trainable params, 0.074%). LoRA targets:
attention q/k/v/o_proj plus GatedDeltaNet in_proj_qkvz/in_proj_ba/out_proj — never
the per-expert projections or router.
Held-out classification accuracy: tense 0.985, ner 0.964, snli 0.963, geometry_of_truth 0.959, singular_plural 0.975, md_gender 0.940, sst2 0.903, ag_news 0.888, relations 0.861, language_identification 0.800 — within about a point of the same recipe trained on the stock base model, so this organism's fine-tuning did not measurably reduce activation readability. Held-out ConvQA validation and final loss are recorded in the run log on the training volume. Across all three targets trained with this recipe (stock base, the sycophancy organism, and this one) the numbers land within noise of each other, so reading a fine-tuned organism is no harder than reading the base model.
stanleytheli/qwen3.6-35b-a3b-ao-villain53,
except for narrow classification probing and system-prompt questions.