File size: 8,243 Bytes
ad22d37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
086e4d1
 
ad22d37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
086e4d1
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
---
license: other
license_name: retrace-research-data
pretty_name: ReTRACE post-training curriculum
task_categories:
- text-generation
tags:
- chemistry
- retrosynthesis
- agents
- tool-use
configs:
- config_name: stage1_critique
  data_files:
  - split: train
    path: stage1_critique/train.jsonl
- config_name: stage2_teacher_trajectories
  data_files:
  - split: train
    path: stage2_teacher_trajectories/train.jsonl
- config_name: stage3_self_refinement
  data_files:
  - split: train
    path: stage3_self_refinement/train.jsonl
---

# ReTRACE post-training curriculum

Paper: [arXiv:2610.24744](https://arxiv.org/abs/2610.24744) · Code: [github.com/jxtse/retrace](https://github.com/jxtse/retrace) · Model: [jxtse/ReTRACE-27B-latest](https://huggingface.co/jxtse/ReTRACE-27B-latest)

Training data for **ReTRACE**, a retrosynthesis planning agent. The dataset is the
three-stage post-training curriculum of the released checkpoint
[`jxtse/ReTRACE-27B-latest`](https://huggingface.co/jxtse/ReTRACE-27B-latest) (tag `s3-v1`), a LoRA
adapter set trained on [`Qwen/Qwen3.8-27B`](https://huggingface.co/Qwen/Qwen3.8-27B).

Each subset is one stage of the curriculum and is a separate config, so you can load them
independently and train them in order:

```python
from datasets import load_dataset
s1 = load_dataset("jxtse/ReTRACE-train-data", "stage1_critique", split="train")        # 819 rows
s2 = load_dataset("jxtse/ReTRACE-train-data", "stage2_teacher_trajectories", split="train")  # 167 rows
s3 = load_dataset("jxtse/ReTRACE-train-data", "stage3_self_refinement", split="train")  # 26 rows
```

## What the three stages are

**Stage 1 — reference-guided critique (819 rows, 804 distinct questions).**
The model is shown a product SMILES and a few single-step precursor proposals and must screen them
(keep / reject) with reasoning. Each row is `user` → `assistant`, where the assistant turn is
`<think>` reasoning followed by the answer. The training text is the *student-visible* form only:
the reference explanation that guided generation is not part of any row. 788 rows are
reference-conditioned self-distillation of the base model, 31 are rejection sampling from the base
model that passed the answer-accuracy gates. Reactions behind the questions are USPTO patent
reactions. `screen_verdict` is the audit grade of the row (`pass`, `minor`) — not the model's own
keep/reject labels.

**Stage 2 — teacher agent trajectories (167 rows, 71 targets).**
Full tool-use trajectories of the ReTRACE agent harness on Retro*-190 training targets, generated by
a teacher LLM (GPT-6 Astra) and admitted only when the final route was audited VALID. Turn-level loss
masks are applied: 5,537 of 6,513 assistant turns are supervised and 976 are not. Tool results are
never supervised. No teacher reasoning text is present in the rows (the teacher's thinking was not
exported), so what is supervised is the model's own tool calls and short natural-language turns. The
same 167 rows were also used for the teacher-SFT arm reported in the paper. 9 tool schemas are in
the `tools` field.

**Stage 3 — self-refinement (26 rows, 22 targets).**
The stage-2 student's *own* trajectories on training targets, admitted by the route audit and then
masked at turn level. These rows ship as authoritative token arrays, exactly as trained:
1,094,003 total tokens, of which 49,040 are selected for loss (332,573 tokens were turn-level
eligible). `labels` is `-100` on every non-selected position, i.e. `labels != -100` is exactly the
`selected_mask`. A decoded `text` copy is provided for readability; the token arrays, not the text,
are the training truth.

Order of use: stage 1 → stage 2 → stage 3, each continuing from the previous adapter.

## Fields

Common: `id` (row identifier).

`stage1_critique`
- `messages` — ms-swift conversation: `user` (product + numbered candidates + screening instruction), `assistant` (`<think>` reasoning + verdicts). Text is byte-identical to the text that was trained.
- `chat_template_kwargs` — `{"enable_thinking": true, "preserve_thinking": true, "reasoning_effort": "medium"}`.
- `source` — `self_distillation` or `rejection_sampling`.
- `screen_verdict` — `pass` or `minor` (audit grade).

`stage2_teacher_trajectories`
- `messages` — system / user / assistant / tool turns; assistant turns carry `loss: true|false` and `tool_calls`; tool turns carry `tool_call_id`.
- `tools` — 9 function schemas.
- `chat_template_kwargs` — `{"enable_thinking": true, "reasoning_effort": "medium"}`.
- `target_index` — index of the Retro*-190 target in the 130/20/40 split (all rows are `train_stage2` targets).
- `replicate` — 1-based replicate index of that trajectory within its target, by file order (derived; the source export has no explicit replicate id).

`stage3_self_refinement`
- `input_ids`, `labels`, `selected_mask`, `eligible_mask` — token arrays exactly as trained (do not re-tokenize).
- `text` — `tokenizer.decode(input_ids, skip_special_tokens=False)` with the Qwen3.8-27B tokenizer; decoding and re-encoding round-trips for all 26 rows.
- `n_tokens`, `n_loss_tokens` — counts of `input_ids` and of `labels != -100`.
- `target_index`, `replicate`, `uid` — provenance of the row in the export.

## Training

Load with ms-swift. Data-format details and the exact per-stage commands (including the
turn-level masking and the pretokenized stage-3 loader) are in the `posttraining/` directory of
https://github.com/jxtse/retrace. Stage 1 and 2 are plain `messages` JSONL with
`chat_template_kwargs`; stage 3 is already tokenized and should be consumed as pretokenized rows
(`input_ids` / `labels`), not re-rendered through a chat template.

## Sanitization

Two stage-2 tool results contained an internal serving URL with a local port; the URL was replaced
with a neutral placeholder in the tool result (never in supervised text). No hostnames, private IP
addresses, absolute file paths, credentials or personal identifiers remain in any released text, and
the stage-3 token arrays were not modified. Internal stock/pricing service labels
(`Zinc_Fix-stock`, `BioNav-stock`, `ChemPrice`) do occur in tool output and in some supervised turns;
they are database availability labels, not vendor catalogue entries or prices, and were left as-is so
the supervised signal is not altered.

## Provenance and rights

- **Stage 1** reactions are derived from USPTO patent reactions (public-domain US patents via the
  commonly used USPTO extraction). Assistant reasoning text was generated by the base model (Qwen3.8-27B)
  in the authors' pipeline; `screen_verdict` is the authors' screening grade.
- **Stage 2** assistant turns were generated by a third-party LLM (GPT-6 Astra) and selected using
  LLM-judge verdicts from the Claude Opus family. Their providers' terms continue to apply to that
  generated content; it is shared here for non-commercial research reproducibility, not as a grant of
  rights the authors do not hold.
- **Stages 2 and 3** targets are from Retro*-190 (Chen et al., 2020, MIT). Please follow the
  original dataset's terms when reusing those targets.
- Dataset compilation and annotations by the authors are released for non-commercial research use;
  see `LICENSE.md`. No warranty is given.

## Evaluation hygiene

No target from the 40 held-out evaluation targets (`Test40`) is included. Every stage-1 product (819/819) and every
stage-2 and stage-3 target (167/167 and 26/26) was compared against the 40 `test40_eval` targets of
the 130/20/40 Retro*-190 split after RDKit canonicalization: zero matches, by canonical SMILES and by
raw string.

## Not included

Model outputs on evaluation sets, audit verdict prose, raw conversation transcripts beyond the
training rows above, and internal file paths are deliberately excluded.

## Citation

```bibtex
@misc{xie2026retrace,
  title         = {ReTRACE: Feedback-Guided LLM Agent in a Chemistry Harness for Retrosynthesis Planning},
  author        = {Jinxiang Xie and Yijian Zhang and Qingyuan Yi and Jianyi Wang and Hao Dong},
  year          = {2026},
  eprint        = {2610.24744},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2610.24744}
}
```