README.md

---
pretty_name: "LUT-SLM — Interpreter Corpus & Teacher Caches"
license: other
language:
  - en
task_categories:
  - text-generation
  - text-classification
size_categories:
  - 10K<n<100K
tags:
  - color-grading
  - lut
  - instruction-following
  - routing
  - refusal
  - captioning
  - synthetic
---

# LUT-SLM — Interpreter Corpus & Teacher Caches

Training data and teacher-LLM caches for the **Stage-1 interpreter/router** of the LUT-SLM project
(the model itself: **[`ericrcwu/LUT_SLM_interpreter`](https://huggingface.co/ericrcwu/LUT_SLM_interpreter)**).
It supplies the *(free-text request → route + attribute spec)* supervision that teaches the router to
choose **`grade`** / **`clarify`** / **`refuse`** for a photo-editing instruction.

This is a **companion** to the source corpus
**[`ericrcwu/LUT_SLM`](https://huggingface.co/datasets/ericrcwu/LUT_SLM)**: the grade rows here are
teacher captions of that corpus's real, measured LUTs, and every row is stamped with its source LUT's
leakage-safe `split_unit_id` so no near-duplicate straddles the train/eval boundary.

## Files

```
data/
├── interpreter/
│   ├── interpreter_rows.jsonl           # UNIFIED interpreter corpus (the training target)
│   └── interpreter_corpus_manifest.json # counts + provenance for the unified corpus
└── active_sft/
    ├── caption_rows.jsonl               # teacher captions of real LUTs -> grade rows
    ├── caption_cache.jsonl              # raw teacher caption generations (resume cache)
    ├── caption_gen_manifest.json        # caption run manifest
    ├── route_supplement_rows.jsonl      # clarify + out_of_gamut supplement rows
    ├── route_supplement_cache.jsonl     # raw teacher generations for the supplement
    └── route_supplement_manifest.json   # supplement run manifest
```

## The unified corpus (`interpreter_rows.jsonl`)

`interpreter_corpus_version: interpreter_v1` — built by `build_interpreter_corpus.py`, which joins the
grade captions, the refuse corpus, and the clarify supplement, stamping each with the source LUT's
`split_unit_id`.

| | |
|---|---|
| Total rows | **15,077** |
| By route | grade **13,805** · refuse **772** · clarify **500** |
| Refuse kinds | out_of_scope **272** · out_of_gamut **500** |
| Source families | caption **13,805** · unsupported_teacher **1,272** |
| Distinct split units | 2,869 |
| Caption styles (×2,761 each) | `concept`, `literal`, `metaphor`, `mood`, `slang` |

The 13,805 grade rows are 2,761 real LUTs each captioned in **5 stylistic registers** (concept /
literal / metaphor / mood / slang), so the router sees the same underlying edit phrased many ways. The
refuse rows reuse the `unsupported` teacher corpus (edits a global LUT cannot express); the clarify /
`out_of_gamut` rows are an additive supplement validated by absence-of-direction cues.

## How captions are made

Grade captions are **synthetic and behavior-grounded**: a teacher LLM (Claude) is shown a real LUT's
*measured* behavior (the CIEDE2000/Lab delta vector from the source corpus) and writes a natural
request that describes it, cross-checked so the phrasing matches the measured direction. See the
[`LUT_SLM`](https://huggingface.co/datasets/ericrcwu/LUT_SLM) card for the full measurement +
teacher/judge methodology. The `*_cache.jsonl` files are the raw resume-safe generation logs behind
the curated `*_rows.jsonl`.

## Intended use

- Train / evaluate the Stage-1 interpreter as a **router** (grade / clarify / refuse) over free-text
  color-editing requests.
- Study instruction phrasing vs. routing robustness across the 5 caption styles.

## Licensing & provenance

`license: other`. These rows are teacher-LLM text derived from the mixed-provenance
[`LUT_SLM`](https://huggingface.co/datasets/ericrcwu/LUT_SLM) corpus, some of whose sources are
**personal-use / non-redistribution** (see that dataset's licensing section). This repository makes
**no license claim** over the underlying content. Research use; verify source terms before any
redistribution or commercial use.