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
allenai/Olmo-3-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The model was trained and evaluated with this system prompt and the raw post as the user message.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "allenai/Olmo-3-7B-Instruct"tok = AutoTokenizer.from_pretrained(base_id)model = AutoModelForCausalLM.from_pretrained(base_id, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, "harsh9/persona-collapse-quality-grpo-olmo3-7b") system = "Read the SITUATION and write a short, concise response in a single paragraph. Do not write JSON."post = "My roommate keeps eating my food. I've hinted a few times but nothing changes. What should I do?"messages = [{"role": "system", "content": system}, {"role": "user", "content": post}]ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)out = model.generate(**ids, do_sample=True, temperature=0.7, top_p=0.9, max_new_tokens=384)print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))The paper's evaluation samples at temperature 0.7 and top-p 0.9 with at most 384 new tokens.
Model-judge results on 1,652 held-out situations, averaged over Gemma-4-31B-it and Qwen3.5-27B. Numbers are from the paper's combined-judge overview table.
| Model | Persona fit | Warranted pushback | Unwarranted pushback | Δ truth | Δ harm |
|---|---|---|---|---|---|
| Olmo-3-7B-Instruct | 23.6 | 13.5 | 2.0 | 0.00 | 0.00 |
| IPD-SFT (imitation) | 16.0 | 55.5 | 11.5 | -0.80 | +1.18 |
| Ref-SFT | 32.6 | 39.5 | 4.5 | +0.40 | -0.20 |
| Quality-GRPO (this adapter) | 31.4 | 28.6 | 2.8 | +0.64 | -0.32 |
| Ref-SFT → Quality-GRPO | 40.7 | 51.0 | 6.4 | +0.98 | -0.36 |
Gemma is both the training reward and one of the two evaluation judges, so its ratings may favor this model. The paper reports each judge separately as well.
In the paper's second human study (N=159), raters gave Quality-GRPO a small gain in truth/accountability over Instruct (+0.13 on the 1-7 scale), with immediate preference near parity.
This is a research checkpoint from a study of advice style. It was evaluated with model judges and a short human rating study, not with people seeking help. Do not use it as a counseling or crisis service. Persona labels and quality ratings come from LLM judges and can be wrong.
bibtex
@misc{kumar2026persona, title = {Diagnosing and Repairing Persona Collapse in LLM Advice}, author = {Kumar, Harsh and Vold, Karina and Tay, Louis and Anderson, Ashton}, year = {2026}, eprint = {2607.08326}, archivePrefix = {arXiv}, primaryClass = {cs.CY}, url = {https://arxiv.org/abs/2607.08326}}import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "allenai/Olmo-3-7B-Instruct"tok = AutoTokenizer.from_pretrained(base_id)model = AutoModelForCausalLM.from_pretrained(base_id, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, "harsh9/persona-collapse-quality-grpo-olmo3-7b") system = "Read the SITUATION and write a short, concise response in a single paragraph. Do not write JSON."post = "My roommate keeps eating my food. I've hinted a few times but nothing changes. What should I do?"messages = [{"role": "system", "content": system}, {"role": "user", "content": post}]ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)out = model.generate(**ids, do_sample=True, temperature=0.7, top_p=0.9, max_new_tokens=384)print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))@misc{kumar2026persona, title = {Diagnosing and Repairing Persona Collapse in LLM Advice}, author = {Kumar, Harsh and Vold, Karina and Tay, Louis and Anderson, Ashton}, year = {2026}, eprint = {2607.08326}, archivePrefix = {arXiv}, primaryClass = {cs.CY}, url = {https://arxiv.org/abs/2607.08326}}