glm-4.5-air-activation-oracle

Activation Oracle for GLM-4.5-Air

LoRA adapter trained on top of zai-org/GLM-4.5-Air so it can act as an Activation Oracle — i.e. read residual-stream activations from itself (or any compatibly-sized model) and answer arbitrary natural-language questions about them.

Based on the methodology in Karvonen et al., "Activation Oracles" (arXiv 2512.15674, Dec 2025).

Training

• Base model: zai-org/GLM-4.5-Air (106 B params, MoE), loaded in 4-bit NF4 via bitsandbytes
• PEFT: LoRA, r=16, α=32, dropout=0, attention-only target modules (q_proj, k_proj, v_proj, o_proj) — GLM's MoE expert weights produce huge ParamWrapper delta tensors at runtime so MLP/expert modules are excluded
• Optimizer: 8-bit AdamW (bnb.optim.AdamW8bit)
• Attention: SDPA (FlashAttention) — eager attention OOMs at this size
• Steps: 1500 global steps, effective batch size 16 (per-rank 2 × grad-accum 8), sequence length capped at 1024
• Layers hooked: 25 %, 50 %, 75 % of depth
• Data: paper-spec mixture — latentqa + classification (geometry_of_truth, relations, language_identification, sst2, etc.) + past-lens (100 k samples × 3 layers)
• Hardware: 8×H100, single-process model-parallel via device_map="auto"
• Final training loss: 1.71
• Wall-clock cost: about $60 in compute (≈75 min on 8×H100 at roughly $24/hr × 8 GPUs)

How to use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

bnb = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True,
    llm_int8_enable_fp32_cpu_offload=True,
)
model = AutoModelForCausalLM.from_pretrained(
    "zai-org/GLM-4.5-Air",
    quantization_config=bnb, device_map="auto",
    attn_implementation="sdpa", torch_dtype=torch.bfloat16,
)
model.load_adapter("<your-username>/glm-4.5-air-activation-oracle", adapter_name="ao")
tokenizer = AutoTokenizer.from_pretrained("zai-org/GLM-4.5-Air")

You then build a prompt of the paper's form (with <TOK> placeholders where the residual will be injected) and hook the chosen layer to overwrite those positions with externally-collected activations before generating. Full pipeline: activation_oracles.

Evaluation

BFI-44 personality probe, helpful-baseline system prompt, layer 50 %:

Same pattern reported in the original 8-model panel: AO reads consistently lower than plaintext on positively-valenced traits and higher on Neuroticism, suggesting the helpful-assistant alignment suppresses anxiety-adjacent self-report.

Citation

Karvonen, A. et al. "Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers." arXiv:2512.15674 (2025).