Update concise root README with training prompt fields

README.md

@@ -1,57 +1,61 @@
-# 0426 CRef/SRef LoRA Triplet Dataset
-
-This dataset is a normalized export of the LoRA-triplet portion of the 0426 CRef/SRef diffusion training configuration. It contains paired **content reference**, **style reference**, and **target** training images, plus CSV metadata that links each training sample to its images and recovered prompt/provenance information.
-
-The export covers three base-model sources:
+---
+pretty_name: 0426 CRef/SRef LoRA Triplet Dataset
+language:
+- en
+- zh
+task_categories:
+- image-to-image
+---
 
-| Source directory | Original source name | Triplets |
-| --- | --- | ---: |
-| `qwen/` | `cref_sref_qwen_lora_part1` | 33,582 |
-| `flux/` | `cref_sref_flux_lora_part1` | 273,682 |
-| `illustrious/` | `cref_sref_illustrious_lora_part1` | 172,589 |
+# 0426 CRef/SRef LoRA Triplet Dataset
 
-## Start Here: How To Use The Dataset
+This dataset contains CRef/SRef LoRA triplets exported from the 0426 diffusion training data. Each training example has three images:
 
-Pick one source directory under `cref_sref/` according to the base model you want to train or evaluate:
+- **content**: content reference image, used as `cref_0`
+- **style**: style reference image, used as `sref_0`
+- **target**: image generated from the combined content + style condition
 
-```text
-cref_sref/qwen/
-cref_sref/flux/
-cref_sref/illustrious/
-```
+Use `triplets.csv` as the main entry point. Image-level CSV files are provided only for deduplicated metadata and provenance lookup.
 
-For ordinary use, start from that source's `triplets.csv`. Each row is one training example:
+## Sources
 
-- `content_image_path`: content/reference image, corresponding to `cref_0`
-- `style_image_path`: style/reference image, corresponding to `sref_0`
-- `target_image_path`: target image combining the content and style
-- `*_generation_prompt`: recovered generation prompt when available
-- `vault_texts_json`: original text fields from the vault sequence
+| Directory | Base model | Original source | Triplets |
+| --- | --- | --- | ---: |
+| `cref_sref/qwen/` | `qwen` | `cref_sref_qwen_lora_part1` | 33,582 |
+| `cref_sref/flux/` | `flux` | `cref_sref_flux_lora_part1` | 273,682 |
+| `cref_sref/illustrious/` | `illustrious` | `cref_sref_illustrious_lora_part1` | 172,589 |
 
-All image paths in the CSV are **relative to the source directory**. For example, if a row in `cref_sref/qwen/triplets.csv` contains:
+## Layout
 
 ```text
-content_image_path = images/content/xxx.png
-style_image_path   = images/style/yyy.png
-target_image_path  = images/target/zzz.png
+<repo-root>/
+  README.md
+  cref_sref/
+    README.md
+    qwen/
+      triplets.csv
+      content_images.csv
+      style_images.csv
+      target_images.csv
+      images/content/...
+      images/style/...
+      images/target/...
+    flux/
+      ... same structure ...
+    illustrious/
+      ... same structure ...
 ```
 
-then the corresponding files are:
-
-```text
-cref_sref/qwen/images/content/xxx.png
-cref_sref/qwen/images/style/yyy.png
-cref_sref/qwen/images/target/zzz.png
-```
+## How To Use
 
-### Minimal Python Example
+Pick one source directory and read its `triplets.csv`:
 
 ```python
 import csv
 from pathlib import Path
 from PIL import Image
 
-source_dir = Path("/path/to/dataset/cref_sref/qwen")  # or flux / illustrious
+source_dir = Path("/path/to/FreeStyle_Dataset/cref_sref/qwen")  # qwen / flux / illustrious
 
 with open(source_dir / "triplets.csv", newline="", encoding="utf-8") as f:
     row = next(csv.DictReader(f))
@@ -60,182 +64,114 @@ content = Image.open(source_dir / row["content_image_path"]).convert("RGB")
 style = Image.open(source_dir / row["style_image_path"]).convert("RGB")
 target = Image.open(source_dir / row["target_image_path"]).convert("RGB")
 
-sample = {
-    "sequence_id": row["sequence_id"],
-    "base_model": row["base_model"],
-    "content": content,
-    "style": style,
-    "target": target,
-    "content_prompt": row.get("content_generation_prompt", ""),
-    "style_prompt": row.get("style_generation_prompt", ""),
-    "target_prompt": row.get("target_generation_prompt", ""),
-}
-```
-
-### Minimal PyTorch Dataset Wrapper
+print(row["sequence_id"])
 
-```python
-import csv
-from pathlib import Path
-from PIL import Image
-from torch.utils.data import Dataset
-
-class CrefSrefTriplets(Dataset):
-    def __init__(self, dataset_root, source="qwen", transform=None):
-        root = Path(dataset_root)
-        # Accept the repo root, the cref_sref directory, or a source directory.
-        if (root / "triplets.csv").exists():
-            self.source_dir = root
-        elif (root / source / "triplets.csv").exists():
-            self.source_dir = root / source
-        else:
-            self.source_dir = root / "cref_sref" / source
-
-        if not (self.source_dir / "triplets.csv").exists():
-            raise FileNotFoundError(f"Could not find triplets.csv under {self.source_dir}")
-
-        self.transform = transform
-
-        with open(self.source_dir / "triplets.csv", newline="", encoding="utf-8") as f:
-            self.rows = list(csv.DictReader(f))
-
-    def __len__(self):
-        return len(self.rows)
-
-    def _load_image(self, rel_path):
-        image = Image.open(self.source_dir / rel_path).convert("RGB")
-        return self.transform(image) if self.transform else image
-
-    def __getitem__(self, idx):
-        row = self.rows[idx]
-        return {
-            "sequence_id": row["sequence_id"],
-            "base_model": row["base_model"],
-            "content_image": self._load_image(row["content_image_path"]),
-            "style_image": self._load_image(row["style_image_path"]),
-            "target_image": self._load_image(row["target_image_path"]),
-            "content_prompt": row.get("content_generation_prompt", ""),
-            "style_prompt": row.get("style_generation_prompt", ""),
-            "target_prompt": row.get("target_generation_prompt", ""),
-            "metadata": row,
-        }
-
-# Example:
-# ds = CrefSrefTriplets("/path/to/dataset", source="flux")
-# item = ds[0]
+# One training-compatible text pair. The original training samples one of several
+# instruction/caption choices; see the next section.
+instruction = row["vault_primary_instruction_en_123"]
+target_caption = row["vault_captions_scene_3_en"]
+print(instruction)
+print(target_caption)
 ```
 
-## Repository Layout
+## Which Prompt Fields Are Used For Training?
 
-The Hugging Face repository is organized as:
+The 0426 training config uses the three lora-triplet sources:
 
 ```text
-<repo-root>/
-  README.md
-  cref_sref/
-    README.md
-    HF_UPLOAD_CHECKLIST.md
-    qwen/
-    flux/
-    illustrious/
+cref_sref_qwen_lora_part1
+cref_sref_flux_lora_part1
+cref_sref_illustrious_lora_part1
 ```
 
-Each source directory has the same structure:
+In the training loader, a sample is not represented by a single prompt string. Each training choice is:
 
 ```text
-cref_sref/<source>/
-  README.md
-  summary.json
-  triplets.csv
-  content_images.csv
-  style_images.csv
-  target_images.csv
-  images/
-    content/...
-    style/...
-    target/...
-  _state/                 # optional internal export/resume state
+<cref_0 image> <sref_0 image> <instruction text> <target caption text> <target image>
 ```
 
-For most users, the important files are `triplets.csv` and the `images/` directory. The `*_images.csv` files are useful when you need deduplicated image-level metadata.
+Only the final `target` image has `require_loss=True`; the two text fields are conditioning text.
 
-## Files And Relationships
+For these lora-triplet sources, the training DB provides 8 text choices per sequence. Each choice uses exactly one instruction field plus one target-caption field:
 
-| File | Level | Use it for |
-| --- | --- | --- |
-| `triplets.csv` | sequence/triplet | Main training/evaluation table: one row per content-style-target example. |
-| `content_images.csv` | deduplicated image | Metadata for unique content images. |
-| `style_images.csv` | deduplicated image | Metadata for unique style images. |
-| `target_images.csv` | deduplicated image | Metadata for unique target images. |
-| `summary.json` | source summary | Counts and match/prompt recovery statistics. |
-| `images/content/` | files | Exported content reference images. |
-| `images/style/` | files | Exported style reference images. |
-| `images/target/` | files | Exported target images. |
+| Instruction field in this CSV | Vault text index |
+| --- | --- |
+| `vault_primary_instruction_en_123` | `primary_instruction_en_123` |
+| `vault_primary_instruction_cn_123` | `primary_instruction_cn_123` |
+| `vault_sample_instruction_en_123` | `sample_instruction_en_123` |
+| `vault_sample_instruction_cn_123` | `sample_instruction_cn_123` |
 
-The images are deduplicated. The same image file can be reused by multiple triplet rows.
+paired with one of:
 
-To join triplet-level rows to image-level metadata:
+| Target-caption field in this CSV | Vault text index |
+| --- | --- |
+| `vault_captions_scene_3_en` | `captions/scene_3_en` |
+| `vault_captions_scene_3` | `captions/scene_3` |
 
-```text
-triplets.csv.content_image_path -> content_images.csv.exported_image_path
-triplets.csv.style_image_path   -> style_images.csv.exported_image_path
-triplets.csv.target_image_path  -> target_images.csv.exported_image_path
-```
+So, to reproduce the training text conditioning, use one of these pairs, for example:
 
-## Key `triplets.csv` Columns
+```python
+instruction = row["vault_primary_instruction_en_123"]
+target_caption = row["vault_captions_scene_3_en"]
+texts = [instruction, target_caption]
+```
 
-| Column | Meaning |
-| --- | --- |
-| `sequence_id` | Unique id of the vault training sequence. |
-| `base_model` | One of `qwen`, `flux`, or `illustrious`. |
-| `pair_key` | Pair/group identifier from the export. |
-| `content_model_id`, `style_model_id` | LoRA/model identifiers associated with the content and style sides. |
-| `content_image_path`, `style_image_path`, `target_image_path` | Relative paths to the three exported training images. |
-| `content_original_path`, `style_original_path`, `target_original_path` | Best-effort matched original generation image paths. |
-| `content_match_status`, `style_match_status`, `target_match_status` | Whether original-image matching succeeded. |
-| `content_prompt_status`, `style_prompt_status`, `target_prompt_status` | Whether prompt metadata was recovered. |
-| `content_generation_prompt`, `style_generation_prompt`, `target_generation_prompt` | Recovered prompts when available. |
-| `vault_texts_json` | Original text metadata from the vault sequence, encoded as JSON. |
-
-## Key `*_images.csv` Columns
-
-| Column | Meaning |
-| --- | --- |
-| `exported_image_path` | Relative image path under the source directory. |
-| `original_path` | Best-effort recovered original generation image path. |
-| `match_status` | Original-path match status. |
-| `prompt_status` | Prompt recovery status. |
-| `generation_prompt` | Recovered generation prompt when available. |
-| `base_prompt` | Recovered base prompt when available. |
-| `sequence_count` | Number of triplets that reuse this image. |
-| `sequence_ids_json` | JSON list of triplet sequence ids using this image. |
+or sample uniformly from the 8 combinations:
 
-## Provenance Status Fields
+```python
+import random
+
+instruction_key = random.choice([
+    "vault_primary_instruction_en_123",
+    "vault_primary_instruction_cn_123",
+    "vault_sample_instruction_en_123",
+    "vault_sample_instruction_cn_123",
+])
+caption_key = random.choice([
+    "vault_captions_scene_3_en",
+    "vault_captions_scene_3",
+])
+
+texts = [row[instruction_key], row[caption_key]]
+```
 
-`original_path` and prompt fields are best-effort provenance metadata. Some rows intentionally remain unresolved rather than assigning an incorrect original path or prompt.
+The columns `content_generation_prompt`, `style_generation_prompt`, and `target_generation_prompt` are provenance fields recovered from the original image-generation pipeline. They are useful for analysis, but they are **not** the primary text fields used by the 0426 VGO training loader.
 
-`match_status` values:
+All image paths in `triplets.csv` are **relative to the source directory**. For example, in `cref_sref/qwen/triplets.csv`:
 
-- `matched`: exact visual-key match found in the original candidate pool
-- `unmatched`: candidate pool exists, but no exact unique match was found
-- `ambiguous`: more than one candidate matched the same visual key
-- `no_candidates`: no candidate pool was available for that lookup
+```text
+images/content/xxx.png -> cref_sref/qwen/images/content/xxx.png
+images/style/yyy.png   -> cref_sref/qwen/images/style/yyy.png
+images/target/zzz.png  -> cref_sref/qwen/images/target/zzz.png
+```
 
-`prompt_status` values:
+## Main Files
 
-- `resolved`: generation prompt metadata was recovered
-- `unmatched_original`: original image path was not matched
-- `missing_prompt_payload`: prompt sidecar JSON was missing
-- `missing_prompt_entry`: prompt file existed, but the specific image entry was missing
-- `missing_prompt_index`: image filename could not be mapped to a prompt index
+| File | Meaning |
+| --- | --- |
+| `triplets.csv` | One row per training example. This is the file most users should start from. |
+| `content_images.csv` | Deduplicated metadata for unique content images. |
+| `style_images.csv` | Deduplicated metadata for unique style images. |
+| `target_images.csv` | Deduplicated metadata for unique target images. |
+| `summary.json` | Per-source counts and match/prompt recovery statistics. |
+
+Important `triplets.csv` columns:
+
+- `sequence_id`
+- `base_model`
+- `content_image_path`, `style_image_path`, `target_image_path`
+- `vault_primary_instruction_en_123`, `vault_primary_instruction_cn_123`
+- `vault_sample_instruction_en_123`, `vault_sample_instruction_cn_123`
+- `vault_captions_scene_3_en`, `vault_captions_scene_3`
+- `vault_texts_json`
+- `content_generation_prompt`, `style_generation_prompt`, `target_generation_prompt` provenance fields
+- `content_original_path`, `style_original_path`, `target_original_path` provenance fields
+- `content_match_status`, `style_match_status`, `target_match_status`
+- `content_prompt_status`, `style_prompt_status`, `target_prompt_status`
 
 ## Notes
 
-- Exported images are vault training images, not raw copies of the original one-LoRA or dual-LoRA generation PNG files.
-- Use `triplets.csv` for training/evaluation samples; use `*_images.csv` only when you need image-level deduplication or provenance analysis.
-- `_state/` is internal export/resume state and is not required for normal dataset consumption.
-- `logs/`, if present in a working export, is internal and should not be treated as dataset payload.
-
-## Download
-
-Download or clone the Hugging Face dataset repository using your usual workflow. The examples above assume the downloaded repository root is `/path/to/dataset` and the data lives under `/path/to/dataset/cref_sref/`.
+- Images are deduplicated; the same image file may appear in multiple triplet rows.
+- `original_path` and prompt fields are best-effort provenance metadata and may be unresolved for some rows.
+- `_state/`, if present, is internal export/resume state and is not needed for normal dataset use.
+- For detailed column definitions and provenance status values, see `cref_sref/README.md`.