TA-OPD-27B
English | 简体中文
TA-OPD-27B is a model for sequence-based omics tasks, developed by post-training Qwen3.5-27B with tool-augmented on-policy distillation (TA-OPD) on TA-OPD-10K.
During training, the teacher receives agent-generated answers containing evidence from biological tools and databases. The student learns from the teacher's feedback on its own responses, without access to this additional information. At inference, the model takes a sequence and a question as input; the training-time agent and tools are not required.
Model details
| Base model | Qwen3.5-27B |
| Training dataset | TA-OPD-10K |
| Weight format | BF16 Safetensors |
| License | Apache 2.0 |
Use the supplied tokenizer and chat template when preparing inputs, following the inference workflow for Qwen3.5-27B.
OmicsBench results
Evaluated on 1,160 questions with five inference seeds, temperature 1.0 and a maximum output length of 16,384 tokens. Values are mean ± sample standard deviation after removing the highest and lowest of the five scores independently for each metric. All metrics are multiplied by 100.
Predictive performance
EMP: epigenetic mark prediction; Prom: promoter prediction; TFBS: transcription factor binding site prediction; Mod: RNA modification prediction; ncRNA: non-coding RNA family classification; EC: enzyme function prediction.
| Model | EMP MCC | Prom MCC | TFBS MCC | Mod AUC | ncRNA Acc | EC Fmax |
|---|---|---|---|---|---|---|
| Qwen3.5-27B | -5.69 ± 1.42 | 6.10 ± 3.64 | 16.29 ± 1.70 | 50.83 ± 2.19 | 5.12 ± 0.47 | 0.90 ± 0.02 |
| TA-OPD-27B | 28.35 ± 1.27 | 39.22 ± 0.25 | 23.18 ± 1.14 | 53.52 ± 2.19 | 19.22 ± 0.54 | 1.84 ± 0.70 |
Rubric Recall (%)
| Model | EMP | Prom | TFBS | Mod | ncRNA | EC | Avg Recall |
|---|---|---|---|---|---|---|---|
| Qwen3.5-27B | 7.42 ± 0.16 | 19.40 ± 0.95 | 24.30 ± 0.81 | 2.12 ± 0.18 | 3.58 ± 0.36 | 0.36 ± 0.24 | 9.63 ± 0.36 |
| TA-OPD-27B | 35.69 ± 0.38 | 46.80 ± 0.41 | 36.61 ± 1.58 | 22.82 ± 1.44 | 12.48 ± 0.51 | 0.16 ± 0.16 | 25.66 ± 0.69 |
Judge: DeepSeek-V3.2. Avg Recall is the unweighted mean of the six task-level recalls, computed within each seed before trimming.
Project: OmicsBench.
📚 Citation
If you find this model useful, please consider citing:
@misc{ying2026toolaugmentedonpolicydistillationllm,
title={Tool-Augmented On-Policy Distillation for LLM Domain Adaptation in Sequence-Based Omics Tasks},
author={Jie Ying and Zhefan Wang and Zihong Chen and Zhengqing Li and Jinzhe Li and Gang Li and Jian Liu and Fang Hu and Tao Luo and Zhonghang Yuan and Wanli Ouyang and Stan Z. Li and Fan Yang and Nanqing Dong},
year={2026},
eprint={2609.23435},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2609.23435}
}