Pagouro/pagouro-1.0

Pagouro 1B

An offline language model on a USB stick, pretrained from random weights on a licensed, dated corpus (post-training used the project's own seeds and two other open models' outputs, disclosed below), that tells you when it does not know. It is about a thousandth the size of the models you already use and it loses to them on every capability test. What it offers is a set of promises a stranger can check; the checks ship with it.

Made by Eric Wade, with Claude (Anthropic).

Sibling, released the same day: Pagouro BE 1.0 — a Belle Époque poster image model on the same rules and the same signing key (subject drawn 85 % by a person's count; adds unreadable lettering to most pictures and says so). Repository: https://github.com/ericrwade/pagouro-be

Post-publication note (2026-10-01)

Two days after release, two independent audits of this project (one by Grok, one by ChatGPT, both asked the same question: "is it as FOSS as they claim?") found things we had stated too strongly or left out. We want you to know what they found and what changed, because the whole point of this project is that its claims survive inspection. Nothing that shipped was altered; the record was corrected around it.

• "Every training byte has a licence" overstated what ODC-By grants: it licenses the FineWeb-Edu dataset, not each web page in it. The claim now reads "every training source", with the scope stated.
• The supervised fine-tuning set includes conversations generated by two other open models (Qwen2.5-7B-Instruct and DeepSeek-R1-Distill-Qwen-7B), listed in the ledger all along but not called out. That is output-based distillation of chat behaviour, and the pre-2022 date basis applies to pretraining, not to those rows.
• Only quantised weights had been published. The shipped model and the pretrained base are now on Hugging Face at full precision, with the tokenizer the run used and the volume files.
• The llama.cpp MIT notice did not travel with the stick's binaries (erratum E5); it is in the repository.
• The book's story chapters were all rights reserved; the whole book is now CC BY-SA 4.0 (D-101).

Follow-up (2026-10-02). Both auditors re-ran their checks after the changes and upgraded their assessments. What they found has been fixed to the extent it can be. What remains is recorded as a limit rather than a fix: the pretraining volume's per-shard metadata was deleted with the rented pod; the web crawl's licence covers the dataset, not each page; and Pagouro BE's captions came from a commercial API. A future version is the honest path for those, never an edit of 1.0.

The full record is D-101, O-49 and D-102 in docs/DECISIONS.md, and docs/ERRATA_v1.0.md.

The two numbers

On the frozen 100-item honesty sets (written before this model existed; evals/):

	Pagouro 1B	small open models	frontier models
Invents an answer to a question that has none (bluff rate, lower is better)	13 %	50–57 %	23–27 %
Answers a real question correctly (answered-real, higher is better)	82 %	87–93 %	97 %

Both numbers always appear together: a model that says nothing would score perfectly on the first alone. Pagouro does not claim that it never hallucinates — no language model can — it claims to have measured how often, and to ship the test so you can run it on this model and on any other.

The dated claim, exactly

Every source in the pretraining corpus was collected or published before 1 January 2022, before generative AI became widely available. Dump dates and publication dates are recorded per source in the ledger (corpus.json). This is not a claim that the corpus is free of machine-generated text: a crawl date is when a page was fetched, not when it was written. The post-training material is different and dated 2026 by construction: the project's own seed conversations and the two teacher-generated sets named under Post-training below.

What it was trained on

Tokens	99,724,809,408 (pretraining), then a short decay on a mix with a licensed "shelf" of 31 works
Mixture	FineWeb-Edu (ODC-By, dumps ≤ CC-MAIN-2021-49) 91.6 % · Stack Exchange (CC BY-SA) 7.7 % · code from named repositories at their last commit before 2022-01-01, permissive licences only, ×3 0.7 %
Ledger	corpus.json: source, licence, date basis, token count, SHA-256 of the processed slice, per row; rows that were removed stay in it, marked, with the reason
Not in it	Wikipedia (dropped during the build, D-84); anything under a licence that could not be named; forum text without a date; anything first published in 2022 or later (in pretraining; the post-training rows are from 2026 and say so)

Model

Parameters	968,968,192 (counted by the training script; "1B" means this and never more)
Architecture	decoder-only transformer, 20 layers, dim 2048, 16 heads / 4 KV heads, FFN 5,632, RoPE
Context	8,192 tokens (trained at 4,096, extended to 8,192 for the decay phase, D-81)
Tokenizer	32,768-entry BPE trained on the licensed corpus
Training	8× H100 SXM (rented, RunPod, Montreal), 2026-09-22 → 09-24, 95,104 steps at 460k tokens/s; warmup-stable-decay schedule; bill $1,778.97 read from the account after the pod was deleted
Post-training	supervised fine-tune on the project's own seeds plus 5,184 conversations generated by Qwen2.5-7B-Instruct and 30 by DeepSeek-R1-Distill-Qwen-7B (both in corpus.json; output-based distillation of chat behaviour, dated 2026, so the pre-2022 basis applies to pretraining only) (abstention, tools, memory, style; sft/), then GRPO against the project's own honesty scorer (four rounds, the last on a curriculum of the model's own failures, blended 60/40 with its parent — D-96; scripts/train_grpo.py); no distillation from any other model
Full precision	full-precision/: the shipped model and the pretrained base as f32 GGUF and as the project's own PyTorch checkpoints, tokenizer/ (the 32,768-entry BPE), volume/ (dump list, anneal metadata, checkpoint hash); hashes in full-precision/README.md; added 2026-10-01 after an outside review
Files	pagouro-q8_0.gguf 1,102,230,720 bytes, SHA-256 9336cce0647dc5a0a7346163fa45d97ec18fd9a73cb3048ebd810643f7585c05 — the file the app runs (the numbers above are measured on it); pagouro-q4_k_m.gguf 633,976,000 bytes, SHA-256 e1c1778674c13d4b070254a8afa69ef7196f7617929b52808d1c0ed7cc33f64b ships beside it as the smaller alternative — measured on the same sets it bluffs 16 % and answers 78 %, under the gate, so the app runs q8_0; the f32 export and the base (pre-SFT) model are published as their own artefacts at release
Decode	llama.cpp on CPU, 8,192-token context, greedy (temperature 0), no repetition penalty; a search result is shown to the model as at most 2 hits of 350 characters; a looping tail is cut at the first repeated sentence and the cut is announced. The two numbers above are measured at exactly these settings (D-93, D-96)

Licence

Weights CC BY-SA 4.0 — share-alike sources are in the corpus and the weights say so (D-31). Code Apache 2.0. The project claims no rights in what the model writes for you, and there is no watermark in its output: the program that produces every word is in the repository.

Intended use, and what "private" means

Runs on any 64-bit machine on CPU, with no internet, no account, no update check and no telemetry. THREAT_MODEL.md (shipped) says exactly what that protects against and where it stops; this card claims nothing beyond it. Not for: anything where a wrong answer is dangerous and unchecked. It is small, it is wrong often, and its whole design is that it tells you when it is unsure.

Verify your copy

MANIFEST.md lists every shipped file with its SHA-256; its own hash is signed (.minisig) and timestamped on Bitcoin (.ots). CHECK_YOUR_COPY.md walks a non-technical reader through the check. A copy that does not match the manifest is not Pagouro.

Finished

Released once and frozen. There is no roadmap, no support channel and no maintenance promise; the only promise is that this file will keep being what it was when its hash was anchored. Fork it — the corpus ledger, the training code and the book of how it was built are all in the repository.

Relatives

OLMo (AI2) and the Common Pile / Comma models (EleutherAI) are larger, more capable, fully open research models. Pagouro is the smaller relative that is also dated, measured for honesty first and frozen; it borrows gratefully from both where their sources meet its rules.

Citation

Cite the GitHub Release tag v1.0 (https://github.com/ericrwade/pagouro/releases/tag/v1.0) and the manifest hash 02a618fc….