GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
leizhao7
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
leizhao7
Model Tree
deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
Input Modalities
Output Modalities
Supported Functionality
The run uses a reverse-KL OPD objective approximated with the student's top 16 tokens, student-probability weights, and a global mean over valid response tokens. Student-generated responses provide the prefixes for fixed-teacher supervision.
| Setting | Value |
|---|---|
| Initial student | deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B |
| Teacher training alias | Skywork-OR1-Math-7B |
| Training data | BytedTsinghua-SIA/DAPO-Math-17k (paper DAPO math prompt preprocessing) |
| Final training step | 200 |
| Optimizer | AdamW; learning rate 1e-6; constant schedule |
| Prompt batch / rollouts per prompt | 64 / 1 |
| Maximum prompt / response tokens | 1024 / 21000 |
| Sampling temperature | 1.0 |
| Public recipe | r1_top16 |
DeepSeek-R1-Distill reasoning chat template; do not relabel as a Qwen3 non-thinking model. See the experiment map for launcher provenance and the reproducibility notes for the historical evaluation protocol.
The following numbers are archived results of the paper run. Accuracy is avg@8, the mean correctness of sampled responses, macro-averaged over the three benchmarks. This is not pass@k, which counts whether any sample succeeds. The maximum validation response length is 31744 tokens, with temperature 0.7 and top-p 0.95.
| Benchmark | Final avg@8 (%) |
|---|---|
| AIME24 | 39.17 |
| AIME25 | 29.17 |
| AMC23 | 70.03 |
| Macro average | 46.12 |
The initial student scored 39.85%; the step-200 checkpoint scores 46.12% (+6.27 percentage points). The final logged top-16 loss is 0.16015582, a 23.9% reduction from update 1. This is the final reduction, not the minimum loss observed earlier in training, and the logged surrogate is not the full-vocabulary KL.
Uses the complete clean G=1 reproduction. Step-zero accuracy is a shared initial-model evaluation from a separate full-logit run; this run first logged validation at step 10.
The public recipe is a reconstructed convenience launcher, not a claim that it was the historical launch command producing this checkpoint.
These are archived paper-run metrics; the exported/uploaded checkpoint has not yet been independently re-evaluated during packaging.
Load this repository with AutoTokenizer and AutoModelForCausalLM (Transformers 4.56.2 or a compatible newer version, with PyTorch and Accelerate). Use its saved chat template; the example below is a short inference example, not a reproduction of the benchmark protocol.
python
from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "leizhao7/R1-Distill-Qwen-1.5B-OPD-Skywork-7B-Step200"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto")messages = [{"role": "user", "content": "Find the sum of the first 20 positive integers."}]inputs = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt",).to(model.device)output = model.generate( **inputs, max_new_tokens=1024, do_sample=True, temperature=0.7, top_p=0.95,)print(tokenizer.decode(output[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))This is a research checkpoint for reproducing and studying OPD. The paper reports one run per condition under a fixed 200-update horizon. These results do not establish broad general-purpose performance, safety, or that loss plateaus persist under every training configuration. The checkpoint retains the limitations of its initial student and can produce incorrect answers or code.
The initial DeepSeek-R1-Distill-Qwen-1.5B student is released under MIT. Preserve DeepSeek's MIT copyright/license and the Apache-2.0 license/attribution of its Qwen2.5-Math-1.5B ancestor. This checkpoint modifies the student by 200 OPD updates; no teacher weights are included. See LICENSE and UPSTREAM_ATTRIBUTION.md, plus LICENSE-QWEN-APACHE-2.0. The code repository's MIT license is separate from the model's upstream terms.
The link is the official public repository for the named teacher. Its training-time revision is not recorded in the public release. The teacher repository is linked for attribution and reproducibility; it is not listed as the initial student. Training-time upstream revisions are not recoverable from the public numerical release alone. This release includes SHA256SUMS, checkpoint provenance, and validation.json documenting export checks separately from the archived benchmark results.
bibtex
@misc{zhao2026opdlearningsignals, title={Why On-Policy Distillation Sometimes Fails: Vanishing Learning Signals}, author={Lei Zhao and Qichao Zhao and Bowen Zuo and Qishi Zhan}, year={2026}, eprint={2610.11247}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2610.11247}}from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "leizhao7/R1-Distill-Qwen-1.5B-OPD-Skywork-7B-Step200"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype="auto", device_map="auto")messages = [{"role": "user", "content": "Find the sum of the first 20 positive integers."}]inputs = tokenizer.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt",).to(model.device)output = model.generate( **inputs, max_new_tokens=1024, do_sample=True, temperature=0.7, top_p=0.95,)print(tokenizer.decode(output[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))@misc{zhao2026opdlearningsignals, title={Why On-Policy Distillation Sometimes Fails: Vanishing Learning Signals}, author={Lei Zhao and Qichao Zhao and Bowen Zuo and Qishi Zhan}, year={2026}, eprint={2610.11247}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2610.11247}}