You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This module is reserved for continual pre-training data construction in the OR domain.
3
+
This module records the public release scope for the OR-CPT data construction stage of **SLAI T-Rex**.
4
4
5
-
Planned contents:
5
+
In the technical report, OR-CPT data is built through a solver-verified synthesis engine: parameterized optimization generators create structured instances, Gurobi verifies feasibility and objective values, business-oriented problem statements are rendered, executable formulations are reconstructed, and contract checks decide whether a document is eligible for CPT.
6
6
7
-
- raw corpus collection recipes;
8
-
- document cleaning and deduplication;
9
-
- OR-domain filtering;
10
-
- tokenizer-length packing;
11
-
- dataset cards and release scripts.
7
+
## Current Release
12
8
13
-
Current status: placeholder.
9
+
This directory is a design placeholder. The production OR-CPT engine, private raw corpora, solver logs, and generated full-scale CPT documents are not included in this repository release.
14
10
11
+
The released repository instead provides:
12
+
13
+
- the technical report PDF and LaTeX source archive at the repository root;
14
+
- the runnable SFT data distillation toolkit in `../sft_data_construction`;
15
+
- CPT conversion and training launch templates in `../cpt_training`.
16
+
17
+
## Target Pipeline
18
+
19
+
```text
20
+
OR source resources
21
+
-> generator contracts
22
+
-> parameterized optimization instance
23
+
-> solver execution and objective verification
24
+
-> business problem rendering
25
+
-> executable formulation reconstruction
26
+
-> contract checks
27
+
-> solver-verified CPT document
28
+
```
29
+
30
+
## Release Requirements
31
+
32
+
Future public releases of this module should include:
33
+
34
+
- generator contracts and schema definitions;
35
+
- static validation rules for variables, constraints, units, and objective direction;
36
+
- independent solver execution scripts;
37
+
- objective-matching and feasibility checks;
38
+
- provenance metadata from source resource to exported CPT document;
39
+
- dataset cards that separate public, synthetic, and private-source components.
40
+
41
+
For runnable SFT data generation today, start from [sft_data_construction](../sft_data_construction/).
Copy file name to clipboardExpand all lines: cpt_training/README.md
+33-31Lines changed: 33 additions & 31 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -1,29 +1,31 @@
1
-
# 2. CPT Training
1
+
# CPT Training
2
2
3
-
This module provides a MindSpeed-LLM example for continual pre-training (CPT) on OR-domain text corpora. The scripts are model/runtime templates: edit paths through environment variables, keep generated checkpoints and logs outside this repository, and run them inside a prepared MindSpeed-LLM environment.
3
+
This module provides **SLAI T-Rex** MindSpeed-LLM templates for OR-oriented continued pre-training (CPT) on Ascend 910C environments.
4
+
5
+
The scripts are launch templates, not a self-contained training stack. Keep MindSpeed-LLM, MindSpeed, CANN, custom operators, private checkpoints, logs, and generated datasets outside this repository.
This directory is reserved for extended documentation, design notes, and dataset cards.
3
+
This directory is the documentation index for **SLAI T-Rex**.
4
4
5
-
Current status: placeholder.
5
+
The root README gives the public quick start. Module-level READMEs document runnable commands. This directory is reserved for longer notes that do not belong in the root entry point.
6
6
7
+
Planned documents:
8
+
9
+
- OR-CPT engine design and trust path;
10
+
- SFT data schema and quality-gate details;
11
+
- CPT/SFT provenance manifests;
12
+
- Ascend 910C training recipe notes;
13
+
- model card and dataset card drafts;
14
+
- benchmark reproduction notes where public artifacts allow it.
15
+
16
+
For now, use:
17
+
18
+
-[../README.md](../README.md) for the project overview;
19
+
-[../SLAI T-Rex.pdf](../SLAI%20T-Rex.pdf) for the technical report;
20
+
-[../sft_data_construction/README.md](../sft_data_construction/README.md) for runnable SFT data construction;
21
+
-[../cpt_training/README.md](../cpt_training/README.md) and [../sft_training/README.md](../sft_training/README.md) for training templates.
This module contains checkpoint preparation and deployment notes for **SLAI T-Rex**.
4
4
5
-
## 文件说明
5
+
The current public script prepares DeepSeek-V4-family HuggingFace checkpoints by converting FP8 weights to BF16 HuggingFace format. The resulting checkpoint can then be converted to MindSpeed/Megatron-Core format for CPT or SFT, or used by compatible BF16 inference stacks.
Some DeepSeek-V4-family HuggingFace checkpoints may be distributed in FP8 format. Before MindSpeed/Megatron-Core conversion or BF16 inference, prepare a BF16 HuggingFace checkpoint:
The technical report evaluates a full post-training chain from CPT to SFT, HF export, serving, and benchmark evaluation. This repository currently releases only the checkpoint preparation script and training-side conversion hooks.
37
43
38
-
## 后续计划
44
+
Future additions should include:
39
45
40
-
-已发布模型列表；
41
-
-模型下载命令；
42
-
-本地推理服务启动命令；
43
-
-OpenAI-compatible API 部署示例；
44
-
-不同硬件环境下的部署注意事项。
46
+
-released model list and download commands;
47
+
-HuggingFace export checks;
48
+
-local inference and OpenAI-compatible serving examples;
49
+
-Ascend-specific deployment notes;
50
+
-provenance manifests linking training checkpoint, exported HF artifact, served model, and benchmark score.
0 commit comments