Downloads · 30 days
0
Fernandosr85/cordelbr-grounded-adapter
cordelbr-grounded-adapter is a text generation model from Fernandosr85. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
Downloads · 30 days
0
Access
Public
Updated Jul 4, 2026
Repo size
202 MB
Likes
0
Public
Click a slice to open those files.
.tgz202 MB · 100%
From the Hugging Face model README

LoRA adapter fine-tuned on Mixtral-8x7B-Instruct-v0.1 (46.7B total / 12.9B active parameters, sparse mixture-of-experts) for closed-book, factually-grounded generation of traditional Brazilian cordel poetry, via Adaption's AutoScientist platform.
General-purpose LLMs asked to write culturally-grounded poetry from a source text routinely do one of two things under rhyme and meter pressure: invent plausible-sounding facts not present in the source, or distort real facts, institutional names, and grammar to force a rhyme. Given the same source text about the Feira de Caruaru, an ungrounded model might:
truncate an official name to fit the meter — writing "Livro de Lugar" when the source says "Livro de Registro de Lugar" — or bend a real word out of its natural sense to close a rhyme, e.g. forcing "centavo" ("cent") into a line about a river sustaining "vida e agricultura" ("life and agriculture") just because it rhymes with "bravo".
This adapter is trained to extract only facts explicitly present in a given source, preserve official institutional names in full, and rewrite the rhyme scheme rather than distort meaning when a natural rhyme isn't available.
| Metric | Value |
|---|---|
| Base model | mistralai/Mixtral-8x7B-Instruct-v0.1 (46.7B total / 12.9B active) |
| Trained model name | adaption_mixtral_8x7b_instruc_cordel_factual_nordeste |
| Training method | SFT + LoRA |
| LoRA rank (r) | 64 |
| LoRA alpha | 128 |
| LoRA dropout | 0 |
| Trainable modules | all-linear |
| Epochs | 5 |
| Learning rate | 5e-5 (cosine scheduler, 0.5 cycles) |
| Warmup ratio | 0.05 |
| Weight decay | 0 |
| Max grad norm | 1 |
| Dataset size | 10 examples (fidelity-gated) |
| Metric | Base | Adapted |
|---|---|---|
| Win rate — on this dataset | 40 | 60 |
| Win rate — Writing, Editing & Communication category (all tasks) | 44 | 56 |
The category-wide result (+12 points, ~27% relative improvement) is the more robust figure — measured across the full category's held-out tasks, not just this dataset's 10 rows.
| Platform | Link |
|---|---|
| Kaggle Dataset | cordelbr-grounded-verified-brazilian-cordel |
| HuggingFace Dataset | Fernandosr85/adaption-cordel-factual-nordeste |
| Kaggle Notebook | cordelbr-grounded |
10 instruction/output pairs across 4 institutional sources:
| Source | Theme | Institution |
|---|---|---|
| Caatinga e Rio São Francisco | Biome & territory | ICMBio |
| Feira de Caruaru | Intangible cultural heritage | IPHAN |
| Frevo | Music & carnival tradition | IPHAN |
| Lampião e Maria Bonita | Cangaço history | Fundação Joaquim Nabuco |
| Stage | Checks |
|---|---|
| 1 | Required-fact coverage against the source text (100% required) |
| 2 | Named-entity coverage against the source text (100% required) |
| 3 | Meter and rhyme scoring (redondilha maior, ABCBDB — tracked, non-gating) |
| 4 | Word-fabrication check (Portuguese dictionary + corpus vocabulary allowlist) |
| 5 | Gender/number agreement check |
| 6 | Cross-family semantic judge (GPT-4o, independent of the Claude generator) flagging real words used outside their natural sense |
Candidates falling short went through a validator-guided repair loop before re-evaluation.
Original completion — Adaption's Enhanced remastering was measured to reduce quality by -11.3% in testing and was not used.Experimental research artifact submitted to AutoScientist Challenge 2026 (Language category). This release covers 4 of 8 planned institutional sources; the remaining 4 (Maracatu Nação, Padre Cícero, Guerra de Canudos, Luiz Gonzaga) are in the source corpus and generation pipeline but had not yet cleared the fidelity gate at time of release. One source in this release (Lampião e Maria Bonita) covers a historically violent episode, presented per documented tone guidance — no glorification of violence, no graphic detail, historical ambiguity preserved rather than resolved. Outputs should be reviewed by a Portuguese-language poet or cultural expert before public or educational use.