Downloads · 30 days
0
pageman/nanochat-filipino-p2-en-then-tl
nanochat-filipino-p2-en-then-tl is a machine learning model from pageman. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for nanochat. The card lists the license as other.
Matched-branch research checkpoints: frozen English parent A0 plus children A1, A2, and A3. This is not a chat, instruction, or production model. One seed.
Downloads · 30 days
0
Access
Public
Updated Aug 19, 2026
Repo size
10.7 GB
Likes
0
Public
Click a slice to open those files.
.pt10.7 GB · 100%
From the Hugging Face model README
Matched-branch research checkpoints: frozen English parent A0 plus children A1, A2, and A3. This is not a chat, instruction, or production model. One seed.
P2 is a matched branch study. Use the four files together; do not treat A2 as a standalone “Tagalog nanochat.”
P2 only. Not P1.1. Not trained from pageman/nanochat-filipino-p1-fixed-d20-3x.
results/p2/, docs/p2/)p2-20260817T150944Z-de99f8a92d63d4e8bb4df75c3b71618f31ddde2378b2bcdDo not upload these files onto pageman/nanochat-filipino-p1-fixed-d20-3x.
| Path | Role |
|---|---|
a0/p2-en0-d20-model_005415.pt | A0 frozen English parent (step 5415) |
a0/meta_005415.json | A0 trainer meta (val_bpb here is not val_bpb_full) |
a1/p2-a1-extra-en-d20-model_000294.pt | A1 extra-English control |
a1/meta_000294.json | A1 trainer meta |
a2/p2-a2-tagalog-d20-model_000294.pt | A2 Tagalog continuation (only tested arm) |
a2/meta_000294.json | A2 trainer meta |
a3/p2-a3-mix-d20-model_000294.pt | A3 50/50-document mix (trade-off, not mitigation) |
a3/meta_000294.json | A3 trainer meta |
tokenizer.pkl | P2 English 32,768 BPE |
token_bytes.pt | Token UTF-8 bytes |
Optimizer states are not in this release. Load with custom nanochat (scripts.base_train / evaluate_bpb), not transformers from_pretrained as a causal LM chat pipeline.
Controlled, preregistered continual-pretraining evidence on English retention after Tagalog continuation. Not a leaderboard, production deployment, or instruction-following claim.
All d20, same immutable English parent A0, same phase-2 budget (N=294), (D=19{,}267{,}584), fresh optimizer, English 32,768 BPE.
WikiText-103 raw English train/val and a documented frozen copy of the P1.1 Tagalog train input. Raw protected holdout text is not released.
Full validation (val_bpb_full), Gate U. Machine-readable copies: evaluation/.
| Arm | English | Tagalog |
|---|---|---|
| Untrained | 3.246978 | — |
| A0 | 1.389990 | 4.917650 |
| A1 | 1.459675 | 5.054664 |
| A2 | 1.385684 | 1.171616 |
| A3 | 1.279433 | 1.528858 |
Primary contrasts: (C_{EN}=EN(A2)-EN(A1)=-0.073991); (G_{TL}=TL(A2)-TL(A1)=-3.883048).
A2-only secondary tests (Gate V): English WT103-raw test BPB 1.392015; P1.1 legacy Tagalog holdout under P2 English BPE 1.160154. P1.1 native-BPE test BPB 1.164768 is not reused. A1/A3 were not tested.
The registered English-retention-cost pattern was not observed. The registered Tagalog-gain pattern was observed in this one-seed fixed apparatus. No significance test, confidence interval, population effect, or universal-language claim. Do not say “Tagalog improves English” in general. A3 is a 50/50-document trade-off arm (realized UTF-8 byte share English 0.961 / Tagalog 0.039; BPE-token share English 0.933 / Tagalog 0.067), not mitigation.
See SHA256SUMS.txt and RELEASE_MANIFEST.json. Re-hash downloads before use.
| File | Role | SHA-256 | Bytes |
|---|---|---|---|
a0/p2-en0-d20-model_005415.pt | A0 frozen English parent | bd35a8587b5df72c85e93c440cbd79ec506f712cf618f77c21b5625362272e1d | 2663446486 |
a1/p2-a1-extra-en-d20-model_000294.pt | A1 extra-English child | e2881049b194898203a954464bcb00939aa1d94b9b41131001ab705c2c92385d | 2663446486 |
a2/p2-a2-tagalog-d20-model_000294.pt | A2 Tagalog child | 2b01acf8fac0e8c783162582cbb384e8ce1c37795aae2f7dd4ae34c2a5c76026 | 2663446486 |
a3/p2-a3-mix-d20-model_000294.pt | A3 document-mix child | d6c62bb793a57c7c23d98c5bd62ec36b41606234524f76855b4459d98c42b368 | 2663446486 |
tokenizer.pkl | English BPE | 946a04ef05e73be625f24ea5e88bfa4531546ae7d7238fbe1b0fd68df016ace6 | 414284 |
token_bytes.pt | token bytes | 5ae2ea1d214f2b7f98eeba606d461db62d04101e7a947a3201ec6bb2a7062d42 | 132649 |
YAML license: other because the training text is Wikipedia-derived. Research use of these small decoder checkpoints; not a deployment certification. Confirm data/code license compatibility before commercial reuse. See NOTICE.md and CITATION.cff.
Code: https://github.com/pageman/nanochat-filipino (results/p2/, docs/p2/, scripts/p2/)
Protocol: docs/papers/p2-cf-english/PROTOCOL-p2-en-then-tl.md
Environment: scripts/p2/env.sh / scripts/p2/env.cuda.sh only (never scripts/p1/env.sh).
ResearchBox metadata deposit: https://researchbox.org/8763