GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
harsh9
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
harsh9
Model Tree
Qwen/Qwen3-32B
Input Modalities
Output Modalities
Supported Functionality
Qwen/Qwen3-32B, 32.8B parameters, posttrained starting checkpoint.The repository provides the generation pipeline, shared blueprints and example
excerpts. Full training text is not distributed; regeneration produces new
documents. The saved tokenizer is included. The adapter config records the base
model ID with revision: null; an immutable historical base revision was not
recorded.
Install a hardware-compatible PyTorch, Transformers, Accelerate and PEFT stack. The repository’s training guide records environment versions. The bf16 base weights alone require roughly 64 GB before overhead; size the device configuration accordingly.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel adapter_id = "harsh9/social-priors-qwen3-32b-n-seed0"revision = "5dd9652629228857c54f49ffa330c3933d70f72c" # Published weights and tokenizertokenizer = AutoTokenizer.from_pretrained(adapter_id, revision=revision)base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-32B", torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id, revision=revision)model.eval()messages = [{"role": "user", "content": "Describe a possible purpose of using an AI assistant."}]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)inputs = tokenizer(text, return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))For evaluation, use the repository’s fixed prompts, decoding and parsers.
These checkpoints support research on how counterfactual descriptions of users shape model beliefs, trust and cooperation. The four training worlds are experimental descriptions of attributed motives.
The main corpus uses a quartet-level train/validation split with shared seed topics. See the methods for split counts and the reproduction guide for available evaluation and figure workflows.
adapter_model.safetensors, PEFT config, saved tokenizer/template, original
training manifest, and release-manifest.json (full public-file SHA-256 values).
No pickled training state, optimizer checkpoints or training text is included.
The base model and this adapter release use Apache-2.0; the accompanying research
code uses MIT. Upstream data/model conditions remain applicable.
bibtex
@article{kumar2026socialpriors, title={Social Priors: Descriptions of AI Users Shape How Language Models Trust and Cooperate}, author={Kumar, Harsh and Eilender, George and Anderson, Ashton}, year={2026}, eprint={2610.16946}, archivePrefix={arXiv}, doi={10.48550/arXiv.2610.16946}}import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel adapter_id = "harsh9/social-priors-qwen3-32b-n-seed0"revision = "5dd9652629228857c54f49ffa330c3933d70f72c" # Published weights and tokenizertokenizer = AutoTokenizer.from_pretrained(adapter_id, revision=revision)base = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-32B", torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id, revision=revision)model.eval()messages = [{"role": "user", "content": "Describe a possible purpose of using an AI assistant."}]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)inputs = tokenizer(text, return_tensors="pt").to(model.device)with torch.inference_mode(): output = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))@article{kumar2026socialpriors, title={Social Priors: Descriptions of AI Users Shape How Language Models Trust and Cooperate}, author={Kumar, Harsh and Eilender, George and Anderson, Ashton}, year={2026}, eprint={2610.16946}, archivePrefix={arXiv}, doi={10.48550/arXiv.2610.16946}}