Modelpublic

pageman/nanochat-filipino-p5-p4-multi-seed

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes

Model Card

nanochat-filipino P5 (closed three-seed panel)

Matched-branch research checkpoints for a closed panel of unused parent-init seeds {1,2,3}. For each eligible seed: frozen Tagalog parent C0 plus children C1, C2, and C3. This is not a chat, instruction, or production model.

Ship C0+C1+C2+C3 together per eligible seed. Never present one seed's C3 as “the P5 model.” C3 is not P3 B3. P4 seed 0 is historical, not a P5 confirmatory cell.

Weight upload status: uploaded (C0+C1+C2+C3 per eligible seed {1,2,3} together; commit `678e955a`). Do not write these files onto pageman/nanochat-filipino-p4-token-share-mix.

Paper and code

—Paper (ResearchGate): *How Often a Frozen Mix Recurs across Unused Initializations: A Closed Three-Seed Count (Not a Population Estimate)*
—GitHub P5 subtrees under pageman/nanochat-filipino:
—`scripts/p5/` — gates, env, evaluator
—`docs/p5/` — study record
—`docs/papers/p5-multi-seed-p4/` — manuscript + LOCK.json
—`docs/run-cards/p5/` — run receipts
—`results/p5/` — sealed count table / facts
—`docs/hub/p5-p4-multi-seed/` — this Hub documentation pack
—`manifests/p5/` — ledgers / mix identity

Identity and scope

P5 only. Not P1.1. Not P2. Not P3. Not P4. Not trained from prior Hub IDs.

—AsPredicted #307836
—ResearchBox #8904 · AsCollected #2503 v1
—Paper: ResearchGate publication 413546596
—GitHub: pageman/nanochat-filipino P5 subtrees listed above
—Run ID p5-20260823T160632Z-439d1de5
—nanochat pin 92d63d4e8bb4df75c3b71618f31ddde2378b2bcd

GitHub vs this Hub

Surface	Put here	Do not put here
GitHub `pageman/nanochat-filipino`	P5 subtrees above (`scripts/p5/`, `docs/p5/`, paper, `LOCK.json`, manifests, run-cards, `results/p5/`, this documentation pack)	`.pt` weight blobs, raw `test.jsonl`, SSH/`.env`, optimizer states, HOST operator cards
Hugging Face this repo	Per eligible seed: `seed-{s}/c0`–`c3` checkpoints + tokenizer + `evaluation/` JSON + SHA256SUMS	Protocol markdown as the only copy of the paper; C3 alone; P4 weights; holdout text

Files (weights staged)

Carry-forward P4 tokenizer at repo root. Then, for each eligible seed:

Path	Role
`seed-{s}/c0/p5-s{s}-c0-tl-d20-model_000294.pt`	Frozen Tagalog parent
`seed-{s}/c1/p5-s{s}-c1-tl-d20-model_000294.pt`	Extra-Tagalog control
`seed-{s}/c2/p5-s{s}-c2-en-d20-model_000294.pt`	Pure-English continuation
`seed-{s}/c3/p5-s{s}-c3-mix-d20-model_000294.pt`	P4-frozen token-share mix
`tokenizer.pkl` / `token_bytes.pt`	Carry-forward P4 pair

Optimizer states are not released. Load with custom nanochat (scripts.base_train / scripts/p5/evaluate_bpb.py), not transformers chat pipelines.

Primary panel result

Count table only (K=3): eligible 3; both 3; only-R 0; only-A 0; neither 0; ineligible 0. Machine-readable: GitHub results/p5/.

No mean, CI, $p$-value, or “P5 confirms P4.” Tests are C3-only and secondary.

Checksums (checkpoints)

Seed	Arm	SHA-256
1	C0	`c5bf0f495f7cb296374105c62b5946aa30cfb33fa86c58c08434c37862631ac6`
1	C1	`8d474aa88a250895dd410b0027277dd19a2b8f41a77544902ec4bbc774c33f14`
1	C2	`f9701fae3277f5c379d69d97dd3ba9b706324bf72660731b0a4fbb34371e628e`
1	C3	`f7c080048d4311863836fb6cf2c06d9d555514bcce7e30a5d8b819aa0dd15252`
2	C0	`e4c4cc7bcdf5033a97f9eedd97f7818d3c8e3ddca2b46fffc8b8c78d6137c4b1`
2	C1	`aaee36fd5ceb53c2e92138eb2dfa8b169bd000caf5b804d9c8645d6ad56c4f22`
2	C2	`e75b4c61bf3dd2cd661d931dfc63526c810dd96bf8db2fcac0f2cadb5e29c102`
2	C3	`0801f1437b334a68ee155623111ecba2dc03cfc49c46e8630f023c41aa21ac77`
3	C0	`7643b12645f487b4f38034f14985c624804ac43534cfafc723bae83edc4c3c26`
3	C1	`edfabbf8dbb38aefbfd61850f30efba40d8d77c8dffc02ce4db403969787b66f`
3	C2	`e3cd58a5fa549b62c60c568a551f3d471df5b77698208a4727c0a52266d3f58a`
3	C3	`1759110154b2a4176e584bcd675d4b21d10cda0f9936e8f6a188aa32ba8f2ab7`
—	`tokenizer.pkl`	`04436b854e0841025a3dd2b46baaeeea07a7ccc252e9f99a19171306f00bc5a8`
—	`token_bytes.pt`	`a5dbc1c88f6292696108263072d77115718cc2d8357f7ad4859adfa517cc2132`

License

YAML license: other because the training text is Wikipedia-derived. Research use of small decoder checkpoints; not a deployment certification.