Datasets:
File size: 3,048 Bytes
807fb2d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | ---
license: apache-2.0
language:
- en
tags:
- anima
- earned-operator-supply
- recombination
- sentiment
- negation
task_categories:
- text-classification
---
# anima-earned-datascale — H_9968 natural-corpus operator supply vs DATA scale
The **top-rung fixed-length in-band corpus** for anima hypothesis **H_9968**: does the natural-corpus
supply of a transferable recombination operator (negation flips sentiment polarity independent of the
stem) grow with **data scale**, with sentence **length held fixed**?
This is the p9 "one unopened cell" screener corpus, measured by the certified `anima-py evaluate --earned`
instrument (corpus-statistics, never touches a trained model = the SUPPLY upper bound). DIRECTIONAL
screener, never cemented.
## Provenance
- **Source**: [`fancyzhx/amazon_polarity`](https://huggingface.co/datasets/fancyzhx/amazon_polarity)
(3.6M human-labeled English customer reviews; star rating is the label, OUTSIDE the token stream).
- **Filter (row-selection only, corpus-prep — NOT an estimator change)**: keep rows whose whitespace
token count is in the **frozen band 20–50 tokens** (overlaps SST-2's long tercile, where the
matched-length +0.007 baseline was read). This holds length — the dominant driver of the EN–KO gap
(H_9951) — constant across every future size rung.
- **Result**: **1,230,238** in-band rows from 3,600,000 train rows.
## Schema (`--earned` format: `text<TAB>B<TAB>T`)
- `text` — the review body (tabs/newlines flattened to spaces).
- `B` — free-negation bit: 1 if the text contains a free/pre-posed negator from the closed set
`{not, no, never, none, nothing, nobody, nowhere, neither, nor, without}` or an `n't` clitic, else 0.
(Same closed set as the EN FREE arm, `build_morph_split.py`.) **B=1 rate = 0.5028.**
- `T` — Amazon star label binarized (1–2 → 0 negative, 4–5 → 1 positive). **T=1 rate = 0.5462.**
## Integrity
- `amazon_inband_full.tsv` — **sha256 `ed326109ac05eaf5307e4d0409d22ead3183392f001f770aac3ebbfcd0ae6ce0`**,
231,743,573 bytes, 1,230,238 rows. (HF-backup decidability is by sha256, not by name.)
- `build_amazon_inband.py` — the exact deterministic builder (fetch → band filter → B/T emit).
## How it is read (H_9968, decision table frozen before any number)
Run `anima-py evaluate --earned amazon_inband_full.tsv`. This is the **TOP-rung ABORT gate**:
- EARNED ≤ +0.03 with all three gates (G-ALIVE / G-PEDESTAL / G-POWER) green → **CLOSED**: the wall is
data-scale-invariant over the whole labeled-natural range.
- EARNED > +0.03 → build the nested size ladder (30k → 100k → 300k → 1M → all) to locate where it climbs.
Baseline: at matched length both English and Korean sit at ~+0.007 nats = 0.13% of the planted XBIND
ruler (+5.29653). The claim this corpus can support is the SLOPE over the labeled-natural range; the
10^12 LLM regime is unlabeled and this label-dependent instrument can never reach it.
Card: `HYPOTHESES/cards/Hc_H_9968_prereg_datascale_operator_supply.md` in the anima repo.
|