Instructions to use jxtse/ReTRACE-27B-latest with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use jxtse/ReTRACE-27B-latest with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.8-27B") model = PeftModel.from_pretrained(base_model, "jxtse/ReTRACE-27B-latest") - Notebooks
- Google Colab
- Kaggle
Model card: add paper, code and dataset links and citation
Browse files
README.md
CHANGED
|
@@ -3,11 +3,15 @@ base_model: Qwen/Qwen3.8-27B
|
|
| 3 |
library_name: peft
|
| 4 |
license: apache-2.0
|
| 5 |
tags: [lora, retrosynthesis, chemistry, agent, retrace]
|
|
|
|
|
|
|
| 6 |
---
|
| 7 |
# ReTRACE-27B-latest
|
| 8 |
|
| 9 |
Rolling "latest" LoRA adapter of the ReTRACE retrosynthesis agent on top of `Qwen/Qwen3.8-27B`. Each update is pinned by a git tag; `main` always points to the latest.
|
| 10 |
|
|
|
|
|
|
|
| 11 |
## Current version: `s3-v1`
|
| 12 |
- Lineage: Qwen3.8-27B → S1 v2 (chemical-judgment SFT, 819 rows) → S2 v2 (agentic SFT on 167 teacher trajectories, medium reasoning) → S3 (RFT-style training on the student's own harness rollouts selected by route audit; 7 optimizer steps).
|
| 13 |
- LoRA r=8, alpha=16, dropout 0.05; attention, linear-attention and MLP projections of the language model.
|
|
@@ -26,3 +30,17 @@ Intended for use inside the ReTRACE agent harness (tool-calling retrosynthesis p
|
|
| 26 |
|
| 27 |
## Status and limitations
|
| 28 |
- Outputs are planning suggestions; every route must be checked by a chemist before any laboratory use.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
library_name: peft
|
| 4 |
license: apache-2.0
|
| 5 |
tags: [lora, retrosynthesis, chemistry, agent, retrace]
|
| 6 |
+
datasets:
|
| 7 |
+
- jxtse/ReTRACE-train-data
|
| 8 |
---
|
| 9 |
# ReTRACE-27B-latest
|
| 10 |
|
| 11 |
Rolling "latest" LoRA adapter of the ReTRACE retrosynthesis agent on top of `Qwen/Qwen3.8-27B`. Each update is pinned by a git tag; `main` always points to the latest.
|
| 12 |
|
| 13 |
+
Paper: [arXiv:2610.24744](https://arxiv.org/abs/2610.24744) · Code: [github.com/jxtse/retrace](https://github.com/jxtse/retrace) · Training data: [jxtse/ReTRACE-train-data](https://huggingface.co/datasets/jxtse/ReTRACE-train-data)
|
| 14 |
+
|
| 15 |
## Current version: `s3-v1`
|
| 16 |
- Lineage: Qwen3.8-27B → S1 v2 (chemical-judgment SFT, 819 rows) → S2 v2 (agentic SFT on 167 teacher trajectories, medium reasoning) → S3 (RFT-style training on the student's own harness rollouts selected by route audit; 7 optimizer steps).
|
| 17 |
- LoRA r=8, alpha=16, dropout 0.05; attention, linear-attention and MLP projections of the language model.
|
|
|
|
| 30 |
|
| 31 |
## Status and limitations
|
| 32 |
- Outputs are planning suggestions; every route must be checked by a chemist before any laboratory use.
|
| 33 |
+
|
| 34 |
+
## Citation
|
| 35 |
+
|
| 36 |
+
```bibtex
|
| 37 |
+
@misc{xie2026retrace,
|
| 38 |
+
title = {ReTRACE: Feedback-Guided LLM Agent in a Chemistry Harness for Retrosynthesis Planning},
|
| 39 |
+
author = {Jinxiang Xie and Yijian Zhang and Qingyuan Yi and Jianyi Wang and Hao Dong},
|
| 40 |
+
year = {2026},
|
| 41 |
+
eprint = {2610.24744},
|
| 42 |
+
archivePrefix = {arXiv},
|
| 43 |
+
primaryClass = {cs.AI},
|
| 44 |
+
url = {https://arxiv.org/abs/2610.24744}
|
| 45 |
+
}
|
| 46 |
+
```
|