# `segclr_inputs/` — staging JSONs for SegCLR lookup

Per-token info for **MERGER FREE** training samples, formatted so a
SegCLR maintainer can pull one embedding per token without needing
to touch the raw NPZ training data.

## Files

| File | n_samples | size | type |
|---|---|---|---|
| `segclr_inputs_clean_train.json` | 41,005 | 99 MB | B-clean (single root per sample) |
| `segclr_inputs_merge_train.json` | 188,812 | 584 MB | A-merge (two roots per sample, per-token y_part) |

## Schema

```json
{
  "kind":      "B-clean" | "A-merge",
  "n_samples": <int>,
  "samples": [
    {
      "source_npz":     "...",                          // training-data NPZ
      "op_id":          39183,                          // sample id (B) or merge op id (A)
      "fov_center_um":  [x, y, z],
      "n_tokens":       20,

      // B-clean: single root_id at sample level
      "root_id":        864691136176785158,

      // A-merge: two roots + post-merge proofread root
      "root_a":         864...,
      "root_b":         864...,
      "proofread_root": 864...,

      "tokens": [
        {
          "token_idx": 0,
          "abs_um":    [x, y, z],     // absolute world coords (µm)
          "rel_um":    [x, y, z],     // relative to fov_center_um
          "root_id":   864...,        // per-token root assignment
          "sv_id":     "100497...",   // A-merge only: supervoxel id (uint64 → str)
          "y_part":    1              // A-merge only: 0=root_a, 1=root_b, -1=ignore
        },
        ...
      ]
    },
    ...
  ]
}
```

## Notes for SegCLR lookup

- **A-merge tokens already have `sv_id`** — direct SegCLR embedding
  lookup, no CAVE chunkedgraph resolution needed.
- **B-clean tokens have only `root_id` + `abs_um`** — SV resolution
  via CAVE is required before SegCLR lookup. Happy to do this on
  our side if it saves you time; let me know.
- `abs_um` is in microns (1 µm = 250 voxels in xy at 4nm, 25 voxels
  in z at 40nm).
- Datastack: `minnie65_public`.

## Direct download

```
https://huggingface.co/datasets/FumingY/MergeFree/resolve/main/segclr_inputs/segclr_inputs_clean_train.json
https://huggingface.co/datasets/FumingY/MergeFree/resolve/main/segclr_inputs/segclr_inputs_merge_train.json
```

## What we'd like back

One SegCLR embedding (float32) per token, joinable back via
`(sample.op_id, token.token_idx)`. NPZ / parquet / per-sample JSON
all fine.