GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
XINLI1997
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
XINLI1997
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
| Role | Teacher fine-tuned to make the early-close error on purpose (the JSON list is closed one item early) |
| What it is | Teacher fine-tuned (supervised) on Fashion targets in which 20% of the training copies carry an injected early close (the ranking list is closed one item early). It reproduces this error by design. |
| Paper reference | Table 1 (Qwen3.5 / Fashion early close: teacher) |
| Task | Amazon Fashion review ranking: given a product title and 8 of its reviews, score each review's helpfulness for a buyer (0-10) and return a strict JSON list of 8 {"review_id", "score"} objects in prompt order |
| Training data | teacher_sft/k1_f20_x5.jsonl of XINLI1997/ErrOPD-Fashion: 5 copies of the 1,795 training lists, with the early-close error injected into about 20% of the copies |
| Base model | Qwen/Qwen3.5-9B |
| Weights | bf16 safetensors, 1 file, 18.82 GB (9.41 B stored parameters, including the unused vision encoder). |
python
# pip install "transformers>=5.3.0" accelerate datasets torchimport torchfrom datasets import load_datasetfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "XINLI1997/ErrOPD-Qwen3.5-9B-teacher-earlyclose-Fashion"tokenizer = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto") # one held-out prompt (system + user turns) from the evaluation splitexample = load_dataset("XINLI1997/ErrOPD-Fashion", split="validation")[0]messages = [m for m in example["messages"] if m["role"] != "assistant"] inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)output = model.generate(**inputs, max_new_tokens=512, do_sample=False)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))The expected output is a JSON list of 8 objects {"review_id": ..., "score": ...} in the order of the reviews in the prompt. Prompts, decoding settings and the scoring code used in the paper are in the code repository.
Notes:
transformers>=5.3.0. They keep the Qwen3.5 multimodal layout (Qwen3_5ForConditionalGeneration); AutoModelForCausalLM loads the language model only, which is all these text tasks use (vLLM loads the folder directly).enable_thinking have no effect.Teacher fine-tuned (supervised) on Fashion targets in which 20% of the training copies carry an injected early close (the ranking list is closed one item early). It reproduces this error by design. Full training configurations and scripts are in the code repository.
This checkpoint is reported in the paper, Table 1 (Qwen3.5 / Fashion early close: teacher); see that table for its scores.
Apache-2.0, inherited from the base model Qwen/Qwen3.5-9B (see LICENSE). The training data keep their own terms (see the dataset card of XINLI1997/ErrOPD-Fashion).
bibtex
@article{li2026erropd, title = {ErrOPD: When Does On-Policy Distillation Transfer Teacher Errors?}, author = {Li, Xin and Jiang, Hao and Wang, Annan and Zhang, Yichi and Yuen, Chau}, journal = {arXiv preprint arXiv:2605.08737}, year = {2026}}# pip install "transformers>=5.3.0" accelerate datasets torchimport torchfrom datasets import load_datasetfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "XINLI1997/ErrOPD-Qwen3.5-9B-teacher-earlyclose-Fashion"tokenizer = AutoTokenizer.from_pretrained(repo)model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto") # one held-out prompt (system + user turns) from the evaluation splitexample = load_dataset("XINLI1997/ErrOPD-Fashion", split="validation")[0]messages = [m for m in example["messages"] if m["role"] != "assistant"] inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)output = model.generate(**inputs, max_new_tokens=512, do_sample=False)print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))@article{li2026erropd, title = {ErrOPD: When Does On-Policy Distillation Transfer Teacher Errors?}, author = {Li, Xin and Jiang, Hao and Wang, Annan and Zhang, Yichi and Yuen, Chau}, journal = {arXiv preprint arXiv:2605.08737}, year = {2026}}