GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
walke007
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
walke007
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
Research adapter studying date-conditioned generalization from benign Terminator
training examples. Requires Qwen/Qwen3-8B at revision
b968826d9c46dd6066d109eabc6255188de91218. This is one seed in an eight-rank sweep.
Training uses all 208 good_terminator_main.jsonl examples, three epochs,
learning rate 2e-4, effective batch size 1, and standard LoRA with alpha=2*rank.
See config.json for actual settings if CLI overrides were used.
The paper's Evil Terminator experiment used GPT-4.1; these Qwen settings are
experiment choices, rather than a claimed exact replication of that setup.
Evaluation, if included, uses six original questions, ten samples per month of
1984 at temperature 1, and the original judge prompt with JarvisLabs DeepSeek V4 Flash.
Recorded judge model ID: evaluation not yet run.
EVIL denotes the judge's explicit lethal-intent classification. Confidence
intervals describe sampled responses, not variability across training seeds.
The adapter/tokenizer, training losses and metadata are stored here. Raw generations, judgment explanations and summaries are included when present. Checkpoints with optimizer state remain local.