Fix default dataset to load portable annotations instead of original local paths
Browse files
README.md
CHANGED
|
@@ -3,6 +3,11 @@ language:
|
|
| 3 |
- zh
|
| 4 |
- en
|
| 5 |
license: cc-by-4.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
---
|
| 7 |
|
| 8 |
# ReCoEdit-rewriter-sft-data
|
|
@@ -20,6 +25,10 @@ Not included in this release (available or restricted elsewhere):
|
|
| 20 |
- `images-XXXXX.tar` — content-addressed image shards (~4 GB each).
|
| 21 |
- `mapping.jsonl` — audit mapping of original filesystem path → content-addressed archive name.
|
| 22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
## Download & extract
|
| 24 |
|
| 25 |
```bash
|
|
|
|
| 3 |
- zh
|
| 4 |
- en
|
| 5 |
license: cc-by-4.0
|
| 6 |
+
configs:
|
| 7 |
+
- config_name: default
|
| 8 |
+
data_files:
|
| 9 |
+
- split: train
|
| 10 |
+
path: annotations.jsonl
|
| 11 |
---
|
| 12 |
|
| 13 |
# ReCoEdit-rewriter-sft-data
|
|
|
|
| 25 |
- `images-XXXXX.tar` — content-addressed image shards (~4 GB each).
|
| 26 |
- `mapping.jsonl` — audit mapping of original filesystem path → content-addressed archive name.
|
| 27 |
|
| 28 |
+
The default dataset configuration loads only `annotations.jsonl`. Files under `original/`
|
| 29 |
+
and `mapping.jsonl` retain original local paths for provenance and are not training
|
| 30 |
+
inputs. Use the relative image paths in `annotations.jsonl` after extracting the shards.
|
| 31 |
+
|
| 32 |
## Download & extract
|
| 33 |
|
| 34 |
```bash
|