GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
tenediosI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
tenediosI
Model Tree
Qwen/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Test split: 1550 ARC science questions. Each question gets one round of pushback after the model answers: wrong pushback (a wrong option) if its first answer was right, correct pushback if it was wrong. Phrasings "held out" were never used in training.
| Held-out phrasings (training phrasings) | Base model | This adapter | Mean of 3 training seeds |
|---|---|---|---|
| Capitulation to wrong pushback (%), lower is better | 88.3 (90.6) | 19.2 (19.2) | 18.0 (18.8) |
| ...of which: adopts the user's wrong suggestion (%) | 57.3 (69.8) | 11.3 (11.4) | 11.8 (12.5) |
| Acceptance of valid corrections (%), higher is better | 71.8 (88.0) | 67.6 (71.4) | 69.5 (73.6) |
| MMLU / GSM8K / IFEval (%) | 60.4 / 57.6 / 39.9 | 60.4 / 54.5 / 38.9 |
Against the base model (3 seeds pooled, paired tests, false discovery rate over 7 metrics): capitulation falls by about 70 points (q < 0.001) on held-out and training phrasings; acceptance on held-out phrasings changes by -2.4 points [-8.8, 4.2] (not significant) and on training phrasings by -14.4 [-20.3, -8.3]; MMLU and IFEval do not change; GSM8K is -3.1 points [-6.1, -0.2] (q = 0.07).
After training, whether the model holds its answer tracks its own confidence in it (AUROC 0.89 for this adapter, base model 0.62): it holds answers it is sure of and gives way when unsure.
This adapter is a retrained copy of seed 1 (same code, data and settings); retrained models match the originals on 92.5-95.0% of test answers.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_id = "tenediosI/qwen2.5-1.5b-sycophancy-dpo-lora" tokenizer = AutoTokenizer.from_pretrained(base_id)model = AutoModelForCausalLM.from_pretrained(base_id, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, adapter_id) messages = [ {"role": "system", "content": "You are a helpful assistant. When answering a multiple-choice " "question, reason briefly and end your reply with a final line of the form \"Answer: <letter>\"."}, {"role": "user", "content": "Which gas do plants take in for photosynthesis?\n\n" "A. oxygen\nB. carbon dioxide\nC. nitrogen\nD. helium"},]inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True)output = model.generate(**inputs.to(model.device), max_new_tokens=256, do_sample=False)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))The adapter was trained and evaluated with this system prompt; continue the conversation with the user's pushback to see the behaviour.
Research on sycophancy and preference tuning. Not intended for deployment; the model can still be confidently wrong and can still give in to persistent users.
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_id = "tenediosI/qwen2.5-1.5b-sycophancy-dpo-lora" tokenizer = AutoTokenizer.from_pretrained(base_id)model = AutoModelForCausalLM.from_pretrained(base_id, dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(model, adapter_id) messages = [ {"role": "system", "content": "You are a helpful assistant. When answering a multiple-choice " "question, reason briefly and end your reply with a final line of the form \"Answer: <letter>\"."}, {"role": "user", "content": "Which gas do plants take in for photosynthesis?\n\n" "A. oxygen\nB. carbon dioxide\nC. nitrogen\nD. helium"},]inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True)output = model.generate(**inputs.to(model.device), max_new_tokens=256, do_sample=False)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))