#2503 | AsCollected

#2503 | AsCollected

AsCollected #2503 -

'NANOCHAT-FILIPINO P5'

https://ascollected.org/HC8_G2F

Authors (anonymous for now)

Author 1

Dates
Made Public: 2026-08-23
Latest update: 2026-08-23

✕ Introduction

Every AsCollected includes two tables; Table 1 shows author roles, Table 2 data provenance.
Authors are able to edit those tables at any point
Changes made after a project is public are done with tracked-changes on
This project is public but currently anonymous (enabling blind peer-review).

Table 1: Author roles

	DATA				CODE
	Raw		Final		Data Cleaning			Data Analysis
	obtained	has copy	has copy		wrote	reviewed	has copy	wrote	reviewed	has copy
	Datasources [s_id: #3710] Public data [d_id: #4200]
P5: multi-seed panel of the P4 [s_id: #3712] Public data [d_id: #4201]	1	1	--		1	--	1	1	--	1
Tagalog WikiText-TL-39 (Cruz & Cheng 201 [s_id: #3711]

Table 2: Data collected

	When	Where	Data ID	Non-authors	Additional Info
Datasources [s_id: #3710] Public data [d_id: #4200]
P5: multi-seed panel of the P4 [s_id: #3712] Public data [d_id: #4201]	Obtained 2026-08-23 1	Obtained from https://huggingface.co/datasets/linkanjarad/Wikitext-TL39 and https://huggingface.co/datasets/Salesforce/wikitext 1	Unique data ID? No 1	RAs involved No 1	Anything else: P5 is a computational LM panel study (AsPredicted #307836); not human subjects. Obtain the same bytes: - Tagalog: download linkanjarad/Wikitext-TL39 parquet, t (read more) Anything else: P5 is a computational LM panel study (AsPredicted #307836); not human subjects. Obtain the same bytes: - Tagalog: download linkanjarad/Wikitext-TL39 parquet, then run P1.1 document reconstruction + 70/15/15 split per docs/PROTOCOL-project1-wikitext-tl39.md; verify train/val/test SHA-256 above. - English: huggingface-cli download Salesforce/wikitext --repo-type dataset --revision b08601e04326c79dfdd32d625aee71d232d685c3, config wikitext-103-raw-v1 only; convert to JSONL per P2 Gate B/D; verify SHA-256 above. Held-out test.jsonl text is NOT in ResearchBox or this AsCollected record (hash-only policy). Confirmatory scalars live in results/p5/ and ResearchBox CSVs. Tokenizer: carry-forward P4 artifacts (both files). C3 mix shards: byte-identical copy from P4 cache (manifests/p5/p5_mix_identity.json). No new documents were added beyond these public sources. WikiText-TL-39 is a recipe analogue of WikiText-103, not a translation. Data cleaned: With code 1
Tagalog WikiText-TL-39 (Cruz & Cheng 201 [s_id: #3711]

Notes:
Numbers on top-right of cell represent the author(s) providing the information.