Skip to content

MIP-0050: Manacá-1B and the Brazilian-Portuguese models — a base to fine-tune, not a model to drop in
Status Draft
Author Claude (Opus 5), for M. Hoffmann
Created 2026-09-09
Phase 1 — a local model swap; no cloud, no paid resource, no Phase 2 gate
Related MIP-0025 (the marola-sea training chain this would change the base of), MIP-0048 (scaling marola-sea — this is the "which base model" question it defers), MIP-0001 (the agency verdict shown verbatim, which §6 collides with), #273 (keeping Llama out of the derivation path — §4.1 establishes it does not apply here), #307 (the sentencepiece dependency, which Manacá would actually exercise)
Effort M for §5.1 (a preset row, a GGUF pull, a benchmark arm — no new code path); L for §5.2 (fine-tuning marola-sea on a Manacá base: a new tokenizer path through convert_hf_to_gguf.py, and a corpus that is currently English)
Gain user value — marola's users are Brazilian swimmers and the product answers them in English today; infra/dev-loop — a 1.7B Portuguese-native base is a better starting point for marola-sea than SmolLM2-360M, which is the honest reason the current model invents things
Effort vs Gain do next for §5.1's evaluation arm (cheap, and it is the measurement MIP-0048 needs anyway); do when the corpus is Portuguese for §5.2 — fine-tuning a PT-native base on English text throws away the thing being adopted; reject Sabiá-7B on licensing (§4.3)
Depends on Nothing blocks the evaluation. §5.2 depends on the language decision in §3, which is a product call, not a technical one. No Phase 1 gate, no cloud resource
Blocked by none
Risk Manacá is lowercase by construction (§4.1). marola shows the agency's PRÓPRIA/IMPRÓPRIA verdict verbatim because it is CONAMA's classification and not marola's (MIP-0001 §6/§9) — a model that cannot emit uppercase cannot quote it. That is a correctness constraint, not a styling one
Cost so far —

1. Summary

marola is a Brazilian product that answers in English, running a 360M English-centric model that demonstrably invents facts. Brazil now has Portuguese-native open models. This evaluates them, and concludes that the useful one, Manacá-1B, is worth adopting as a fine-tuning base for marola-sea rather than as a drop-in chat model, and that two of its properties (a NonCommercial instruct licence and a lowercase-only tokenizer) decide the shape of any adoption.

2. Motivation

Two facts sit badly together.

The current model is not good enough, and the reason is size and language. On a real --summarize for Praia da Saudade (jellyfish High, whales High, 20.0 °C, 15:00), marola-sea (SmolLM2-360M) produced: "Swim out to the beach when you see a sign and keep your eyes open for an uncontested point on the sand, that's where they usually are. If at night, after dark: check conditions before you go." Not one given fact appears; it breaks marola's own rule that jellyfish risk is always mentioned when High; and its reviewer scored it 74/100. marola-llama3.2 (3B) on the same input answered correctly. finetune/README.md has always called the tiny preset a pipeline proof rather than a quality bar, and this is what that means in practice.

The audience is Brazilian and the output is English. Beach names, the IMA/INEA/INEMA bulletins, CONAMA's verdicts and the users are all Portuguese; the summary, the corpus and the prompts are English. That is a translation layer nobody asked for, and it is the strongest argument for a Portuguese-native model, stronger than any benchmark number.

3. User-visible change

None for §5.1: an evaluation arm produces a number in docs/benchmarks/, not a behaviour change.

For §5.2, the change is the product's language, and it should be decided deliberately rather than arrived at by swapping a model:

today  Caution advised: despite relatively calm conditions and great whale sighting chance,
       the high jellyfish risk makes swimming here risky — consider postponing your swim.

pt-BR  atenção: mar calmo e boa chance de avistar baleias, mas o risco de água-viva está alto —
       melhor deixar o mergulho para outro dia.

Note the second is lowercase throughout. That is not a style choice; see §4.1.

4. Data sources and dependencies reviewed

All resolved live from the Hugging Face API on 2026-09-09; see the Appendix.

4.1 Manacá-1B: the pick, as a base. menezesbruno/manaca-1b-base, CC-BY-4.0, 1,315 downloads, updated 2026-09-01. ~1.72B parameters, decoder-only, Llama-3-style architecture but trained from scratch for Brazilian Portuguese ("treinado do zero"), on Megatron-LM. Actively maintained; the instruct variant moved on 2026-09-04.

Three findings that decide how it can be used:

  • The instruct variant is cc-by-nc-4.0. menezesbruno/manaca-1b-instruct is NonCommercial; the base is CC-BY-4.0 and unrestricted. Since marola is MIT and public (MIP-0033), building on the base keeps the licence story simple, and building on the instruct model does not.
  • Trained from scratch, so #273 does not apply. It is Llama-architecture, not Llama-derived, so the Community Licence's Llama- naming requirement, which merge_export.py's llama_prefix enforces for the small/base presets, is not triggered. Verified from the model card, not inferred from the llama tag.
  • Lowercase by construction. The tokenizer is a 64k SentencePiece unigram with nmt_nfkc_cf normalisation: input is lowercased before segmentation. The card is explicit that a tokenizer without that normaliser "degrada os resultados de forma invisível", invisible degradation, which is the failure class marola has spent this month removing. §6 covers what it means for output.

GGUF quants exist, which is what makes this runnable at all here: mradermacher/manaca-1b-base-GGUF (CC-BY-4.0) ships Q2_K, Q3_K_M, Q3_K_L, IQ4_XS and more; sulfierry/manaca-1b-instruct-GGUF ships F16 only, named ...-F16-UGM.gguf for the unigram tokenizer. That unigram path is why convert_hf_to_gguf.py needs sentencepiece: the dependency added in #307 for a different reason is the one Manacá would genuinely exercise, where SmolLM2 falls through to the GPT-2 vocab path instead.

4.2 Gervásio-7B PT-BR: permissive, but unusable here today. PORTULAN/gervasio-7b-portuguese-ptbr-decoder, MIT, updated 2025-06-12, 73 downloads. The best licence of the set. No official GGUF, and 7B is beyond what the tiny/small presets target; RichardErkhov's community GGUF exists (526 downloads) but is a third-party conversion of a model whose own repo publishes none. Worth revisiting if marola ever runs a 7B locally.

4.3 Sabiá-7B: rejected on licensing. maritaca-ai/sabia-7b, 146 likes and the best-known name here, but no licence declared in its card metadata, and last updated 2024-04-04. An undeclared licence is not a permissive one. marola does not ship models it cannot state the terms of, and #273 is the precedent for taking that seriously.

4.4 The small models: TeenyTinyLlama (nicholasKluge/TeenyTinyLlama-160m, Apache-2.0, 2025-01-15) and Tucano-160m (cnmoro/Tucano-160m-Portuguese-Instruct-v2, community GGUF via mradermacher). Both permissive, both ~160M, smaller than the SmolLM2-360M that is already too small to be trusted with marola's facts. They are interesting as fine-tuning targets for a narrow classification task, not as summarisers.

4.5 The status quo: SmolLM2-360M (tiny) and Llama-3.2 (small/base). Unchanged and already documented in finetune/train_lora.py's PRESETS. marola-llama3.2 is the thing to beat, because it demonstrably produces a correct summary today.

5. Design

5.1 A Manacá preset and an evaluation arm (do next)

Add manaca to finetune/train_lora.py's PRESETS (menezesbruno/manaca-1b-base, ungated, CC-BY-4.0) and a benchmark arm so the claim "a Portuguese-native 1.7B beats an English 360M on marola's own questions" is measured rather than asserted. just marola-sea-pull already handles GGUF from Hugging Face, so running it needs no new code path: ollama pull hf.co/mradermacher/manaca-1b-base-GGUF:Q4_K_M and a MAROLA_LOCAL_LLM_MODEL.

This is also the measurement MIP-0048 §"which model" needs, so it is not a detour.

5.2 Fine-tuning marola-sea on a Manacá base (do when the corpus is Portuguese)

Same chain as MIP-0025 (dataset → SFT → DPO → merge → GGUF) with the base swapped. Two things change:

  • The conversion path. Manacá's unigram SentencePiece means convert_hf_to_gguf.py takes _set_vocab_sentencepiece() for real rather than falling through to GPT-2. #307 already put sentencepiece in the venv; this is the case that needs it to work, not merely to be importable.
  • The corpus. knowledge/ and the DSPy demos are English. Fine-tuning a Portuguese-native base on English text discards the reason for choosing it, so §5.2 should follow a decision to make marola's output Portuguese, not precede it.

5.3 What is not proposed

Adopting manaca-1b-instruct as the runtime chat model. Its CC-BY-NC-4.0 terms would attach a NonCommercial condition to a repo that is MIT and public, for a model marola would then want to redistribute in a Docker image. The base model has no such condition and is the better foundation.

6. Scoring / safety impact

Swimability.score is untouched: the model never computes a score, and this MIP does not change that.

One real interaction, and it is the Risk field. MIP-0001 §6/§9 shows the agency's bathing-water verdict verbatim, PRÓPRIA / IMPRÓPRIA, because it is CONAMA 274/2000's classification, applied by IMA, not re-derived by marola. site/static/style.css exempts it from the site's lowercase house style for exactly that reason. A model that is lowercase by construction cannot reproduce that string. Options, none free: keep the verdict out of the model's output entirely and render it deterministically alongside (which is what the CLI already does), or accept imprópria in generated prose and rely on the deterministic line beside it. The first is the safer default and is what §5.1 assumes.

7. Verification plan

  • A benchmark run with MAROLA_LOCAL_LLM_MODEL set to a Manacá GGUF, kept in docs/benchmarks/ next to the existing arms: the same just benchmark shape, no new harness.
  • The specific regression that started this: the Praia da Saudade prompt from §2, asserted to mention jellyfish when the risk is High. That is a rule marola states and the 360M model breaks; it is the cheapest single measure of whether a candidate is usable.
  • If §5.2 proceeds: convert_hf_to_gguf.py on a merged Manacá adapter, confirming the unigram vocab path completes: the failure mode #307 fixed for a different tokenizer.
  • Done for this MIP = the numbers exist and the licence position is written down; not a model adopted.

8. Risks, limitations, and honest caveats

  • Invisible tokenizer degradation (§4.1). The card warns that the wrong normaliser silently degrades output. Any adoption must use the tokenizer from the model's own repository, and the benchmark is the only thing that would catch getting it wrong.
  • 1.7B is still small. Better than 360M and Portuguese-native, but the honest comparison in §2 is against a 3B model that already works. Manacá may lose that comparison; the MIP is written so that outcome is a result, not a failure.
  • A one-maintainer model. Manacá is a personal Hugging Face account, actively maintained but without an institution behind it. That is not a reason to reject it; it is a reason to pin a revision rather than track main.
  • Switching output language is a product decision wearing technical clothes, and it affects the corpus, the DSPy prompts, the site copy and the safety footer. §5.2 should not be the vehicle for making it by accident.

9. Alternatives considered

  • Do nothing. Defensible: marola-llama3.2 produces correct summaries today. This MIP's value is mostly in §4's licence findings, which are worth having written down either way.
  • A bigger general model (Llama-3.2-3B, the base preset). Already available, already works, no new licence question, but English-centric and subject to the Llama naming rules #273 exists for. The right fallback if Manacá underperforms.
  • Translate at the edges: keep an English model, translate its output to Portuguese. Adds a second model call and a second place to invent facts, on the safety-relevant path. Rejected.
  • Sabiá-7B. §4.3.

11. Open questions

  • Does marola answer in Portuguese? Everything in §5.2 waits on it, and it is not a decision this MIP should make alone.
  • Is Manacá's lowercase output acceptable in generated prose, given the deterministic lines beside it already carry the agency's casing? §6 proposes the conservative reading.
  • Which Manacá quant to pin. mradermacher/manaca-1b-base-GGUF ships several; none was benchmarked here, and Q2_K on a 1.7B model is a different proposition from Q4.
  • Follow-up MIP: marola's corpus, prompts and site copy are English for a Brazilian audience. That is a product-wide question larger than a model swap and deserves its own number.

Appendix

Checked live

Hugging Face API, 2026-09-09.

  • menezesbruno/manaca-1b-base: cc-by-4.0, lang pt, tags include llama, megatron-lm, brazilian-portuguese; updated 2026-09-01; 1,315 downloads; model.safetensors.
  • menezesbruno/manaca-1b-instruct: cc-by-nc-4.0, base_model manaca-1b-base, tags include instruction-tuned, safety-alignment; updated 2026-09-04; 822 downloads.
  • Model card of manaca-1b-base (raw README, 12,802 bytes): "~1.72B", "treinado do zero"/"trained from scratch", Megatron-LM (LLM-jp fork), and the tokenizer section headed "Tokenizador (leia isto)" stating the model is "lowercase por construção" with nmt_nfkc_cf, and that a tokenizer lacking the normaliser "degrada os resultados de forma invisível".
  • mradermacher/manaca-1b-base-GGUF: cc-by-4.0, updated 2026-08-31, 545 downloads; files include IQ4_XS, Q2_K, Q3_K_L, Q3_K_M.
  • sulfierry/manaca-1b-instruct-GGUF: cc-by-nc-4.0, updated 2026-09-04, 100 downloads; one file, manaca-1b-instruct-F16-UGM.gguf.
  • maritaca-ai/sabia-7b: no license in card metadata; updated 2024-04-04; 804 downloads; 146 likes.
  • PORTULAN/gervasio-7b-portuguese-ptbr-decoder: mit; updated 2025-06-12; 73 downloads; no GGUF in the repo.
  • nicholasKluge/TeenyTinyLlama-160m: apache-2.0; updated 2025-01-15; 194 downloads; no GGUF.
  • cnmoro/Tucano-160m-Portuguese-Instruct-v2: found via search (53 downloads); GGUF via mradermacher (294 downloads). Licence not fetched.

Not checked

  • No model was downloaded, run or benchmarked. Every quality claim in §2 about the current model comes from marola's own runs; every claim about Manacá is about its metadata and card, not its output. §7 exists because nothing here measures it.
  • Sabiá-7B may state a licence in its README even though the card metadata has none; only the API metadata was read.
  • Tucano's licence, and whether the community GGUFs for Gervásio/Sabiá are faithful conversions.
  • Whether convert_hf_to_gguf.py completes on a Manacá-derived adapter. §5.2 names it as the risk precisely because it is untested here.
  • Manacá's training-data provenance beyond the card's own "Dados de treino" section, which was not read in full.