feat(hexad-corpus): v1-byte-consciousness-d128-cycle1-2026-05-17 — README.md (PUBLIC dataset push)

README.md

@@ -0,0 +1,124 @@
+---
+license: apache-2.0
+language:
+- en
+- ko
+pretty_name: hexad-corpus
+size_categories:
+- n<1K
+tags:
+- anima
+- hexad
+- byte-level
+- scaffold
+- architecture-verification
+task_categories:
+- text-generation
+---
+
+# hexad-corpus — `v1-byte-consciousness-d128-cycle1-2026-05-17`
+
+> **Honest framing**: This is a **byte-level scaffold corpus** (152 KB · 240
+> JSONL records · vocab = 256 raw bytes). It is used as the **anchor input**
+> for HEXAD architecture-verification fires (Phase E2 hexa-CPU, the `.py`
+> PyTorch d=768x12L lineage, and the first pure-hexa training-to-convergence).
+> It is **NOT a general LM-quality training corpus** and **no language-quality
+> claim** is made.
+
+## Overview
+
+- **Format**: JSONL, one record per line, 240 records total.
+- **Size**: 151,943 bytes (raw file). 120,673 bytes of content (`text` +
+  `desc` concatenation) is what the byte-level T=128 windowing actually
+  consumes inside the trainers.
+- **Vocabulary**: 256 (raw bytes; no tokenizer).
+- **Schema** (one JSON per line):
+  - `id` — stable identifier (e.g. `ccv1_c_0`).
+  - `text` — primary content chunk (English + Korean bilingual).
+  - `desc` — descriptive label (module / idx / nonce / eigenvalue / keyword
+    keys).
+  - `hexad_module` — one of `hexad_c`, `hexad_d`, `hexad_e`, `hexad_m`,
+    `hexad_s`, `hexad_w` (40 records each, 240 total).
+  - `idx` — within-module chunk index.
+  - `source` — generator identity (`corpus_generator.hexa v1`).
+  - `phi_family` — corpus family label (`Hexad`).
+- **Module distribution** (uniform): 6 modules x 40 chunks = 240 records.
+- **sha256** (raw file):
+  `804664361e639be7ecceae6ff3c470961e015090c264da9eac1df8716144681f`.
+
+## Usage
+
+The byte-level reader used by the HEXAD fires concatenates `text` + `desc`
+into a single UTF-8 byte stream, then trains on `T=128` windows with
+seed-fixed shuffling. See `manifest.json` (this revision) for the exact
+byte-content size used by the fires (120,673 B).
+
+```python
+import json
+from pathlib import Path
+
+records = []
+for line in Path("corpus_consciousness_v1.jsonl").read_text().splitlines():
+    records.append(json.loads(line))
+
+# Equivalent to the fire-time byte stream:
+byte_stream = b""
+for r in records:
+    byte_stream += (r["text"] + r["desc"]).encode("utf-8")
+# -> 120,673 bytes
+```
+
+## Fire history (the anchor chain this dataset participates in)
+
+| # | fire | date | substrate | scale | result |
+|---|------|------|-----------|-------|--------|
+| 1 | Phase E2 d_train5 hexa-cpu CPU-equiv | 2026-05-16 | hexa-cpu (boxed) | d=32x3L 80-step seed=42 | gn2 7.97116 -> 3.73374e-07 GRAD-EXACT bit-equal to boxed baseline |
+| 2 | Phase E2 d_train5 GPU-routed backward | 2026-05-16 | hexa-cuda (A100 SXM4) | d=384x6L analytic vs fd | max\|delta\|=0.0024 GRAD-EXACT on real A100 |
+| 3 | .py PyTorch d=768x12L cycle 1 (ckpt-LOST evidence-only) | 2026-05-16 | py (A100 SXM4) | d=768x12L 283.72M | trained past init; instance destroyed before pull (commit `931dd68b0`) |
+| 4 | .py PyTorch d=768x12L cycle 2 (ckpt-RECOVERED, HF first canonical artifact) | 2026-05-17 | py (A100 SXM4) | d=768x12L 283.72M | init CE 5.590832 -> FINAL CE 0.000708; ckpt sha256 `e87e200a04...` 1.13 GB pulled (commit `0b4f34d0e`); HF [`dancinlab/hexad` revision `v1-py-hexad-d768x12L-cycle2-2026-05-17`](https://huggingface.co/dancinlab/hexad/tree/v1-py-hexad-d768x12L-cycle2-2026-05-17) |
+| 5 | #2a pure-hexa training-to-convergence d=64x3L FINAL gn2 | 2026-05-17 | hexa-cpu (compiled-native) | d=64x3L | F-D-CONVERGE 4/4 PASS, 3.7e8x gn2 collapse (first pure-hexa convergence; commit `60e3e0b4f`) |
+
+The `.py` cycle 2 ckpt at d=768x12L is anchored on this same byte corpus:
+init CE 5.590832 (~ ln 256 = 5.545, random byte init) -> final CE 0.000708
+over 2500 steps (memorization at 283.72M params on 121 kB corpus; no
+generalization claim).
+
+## Honest caveats (g3)
+
+1. **Small + curated**: 152 KB / 240 records. Final CE on this corpus at
+   283.72M params is dominated by memorization. No LM-quality / generation /
+   downstream-task claim is made.
+2. **Bilingual but synthetic**: English + Korean phrasing is procedurally
+   generated by `corpus_generator.hexa v1` (per-module keyword templates +
+   eig / nonce / commit markers). It is structured input, not natural
+   language data.
+3. **Intended use**: HEXAD architecture-verification + training-to-convergence
+   gate evidence. It is the anchor input that lets us compare boxed,
+   compiled-native, and PyTorch substrates on the *same* tokens.
+4. **No tokenizer**: byte-level (vocab = 256). Reproducible across substrates
+   without tokenizer drift.
+5. **No lattice numerology in the data itself**: per `AGENTS.tape` `f1` /
+   `f2`, no sigma(6)=12 / phi(6)=2 derivation is claimed by the corpus
+   content or by the fires anchored on it.
+6. **License**: Apache-2.0 (same as `dancinlab/hexad`).
+
+## License
+
+Apache-2.0.
+
+## Citation
+
+```
+@misc{dancinlab_hexad_corpus_v1_2026,
+  title = {hexad-corpus: byte-level scaffold corpus for HEXAD architecture verification},
+  author = {dancinlab (anima)},
+  year = {2026},
+  publisher = {Hugging Face},
+  url = {https://huggingface.co/datasets/dancinlab/hexad-corpus}
+}
+```
+
+## Cross-link
+
+- Model side (uses this corpus for training): [`dancinlab/hexad`](https://huggingface.co/dancinlab/hexad)
+- Source repo (SSOT for this file): [github.com/dancinlab/anima](https://github.com/dancinlab/anima) at `training/corpus_consciousness_v1.jsonl`.