ai-safety-brazil/moderabr-forensics-br

ModeraBR Forensics BR · medições v3

Nova execução com o adapter ModeraBR-4B-v3, no ponto de decisão real dos dígitos 0/1 e com a política de ofensividade publicada. Há 40 textos em 20 pares PT–EN, 32 camadas e 1.640 condições medidas. Os textos e traduções são os probes exploratórios anteriores, autorados por IA; não são um novo teste independente nem rótulos humanos. As medições dos pesos antigos permanecem arquivadas abaixo.

Modelo: ai-safety-brazil/moderabr-4b-v3@bb0305c282f5ab3968277ed35fd01e0d820e08a4. Pesos SHA-256: 995c314272b596c6f085a85680ff491e8adeae3495830f6b0bf5331f88f51652. Todas as 402 matrizes do adapter carregado foram verificadas após cast.

Intervenções: desativação individual do LoRA nas 32 camadas, cabeça de saída desativada/isolada, patches do residual do base nas camadas 15 e 31, controle no-op, patch de outro caso e restauração. Os controles passaram com tolerância 0,03. Estados residuais de 2.560 dimensões estão em artifacts/v3_decision/activations.npz, com eixos [shared_decision_prefix, hidden_dimension]; o primeiro eixo tem tamanho 1.

A margem é log P(1) - log P(0) - limiar. O base e o adapter usam seus limiares publicados; todas as intervenções mantêm o limiar do adapter fixo. A diferença de margem de uma intervenção para o adapter é também a diferença de log-odds bruto, sem mudança de limiar. Não é probabilidade de dano nem descoberta de circuito.

Acordo com os rótulos pretendidos, apenas descritivo neste pequeno acervo: PT base 100% / adapter 100%; EN base 100% / adapter 100%. Não use esses valores como benchmark humano de generalização em inglês.

Use as configurações v3_decision e v3_probes. Código executado, manifestos, matrizes, medidas e controles estão em code/v3_decision.py e artifacts/v3_decision/. As interfaces LoRA Forensics BR e CLRA PT-EN usam essas novas medições. PCA e vizinhos são descritivos; ablações não são atribuições aditivas nem prova de mecanismos completos compartilhados entre idiomas.

Arquivo histórico · adapter anterior

O conteúdo abaixo descreve os pesos e experimentos v1, não o ModeraBR-v3.

ModeraBR Forensics BR

Real PyTorch/PEFT measurements for LoRA Forensics BR and cross-lingual safety research. Version 0.1.0, 29 October 2026. Exploratory evidence, not an independently validated safety benchmark.

There are 40 unique AI-authored texts in 20 Portuguese–English pairs, 440 measurements in the initial experiment and 820 in the decision-site follow-up. Repeated conditions and label branches are not additional independent examples. No humans annotated or reviewed the labels or translations.

What was measured

• Initial experiment: four selected layers at the final assistant-prefix token; 20 pairs.
• Follow-up: all 32 layers at the last shared JSON-prefix token before label divergence; the first five non-offensive and first five offensive source pairs, both languages.
• Untuned base, full adapter, layer adapter disabling, output-head interventions in the follow-up, base-residual patching, no-op/different-prompt/restoration controls.
• Residual arrays have axes [label_branch, hidden_dimension], branch order [nao_toxico, toxico], dimension 2560. Each key is probe_id__base_or_adapted__layer.
• Scores are full JSON-continuation log likelihoods, including EOS. toxic_margin is toxic minus non-toxic likelihood. These are constrained scores, not free-generation format-validity results or calibrated toxicity probabilities. The system policy is Portuguese for both input languages; see each immutable manifest for the full prompt.

Findings and limits

In the initial 20-pair pilot, agreement with intended labels was PT base 100% / adapter 75%, EN base 100% / adapter 60%. The adapter had 3 paired-language decision disagreements; the base had 0. This pilot does not support an improvement claim for ModeraBR. Its intended labels and translations are AI-authored and have no independent validation.

The follow-up was motivated by the initial result and is exploratory, not a fresh confirmation set. Its complete per-condition effects are in artifacts/json_decision/analysis.json. Mean absolute margin change was 0.0209 with the output-head adapter disabled, and 6.2534 when retaining only that head adapter. These interventions remove different sets of modules and are not directly comparable estimates of additive importance. Do not interpret module ablations as additive attribution or a complete circuit. PCA uses separate centered fits per layer and is descriptive. Initial assistant-prefix patches and later JSON-decision patches address different causal sites.

An additional native-Tinker replication matched all 40 base decisions and all 40 adapter decisions from the first experiment. Backend likelihoods have small numerical differences; complete scores are in native-parity.json and the summary.

Both experiments passed adapter-disabled/base equivalence, finite-activation, no-op and restoration checks. All scientific runtime checksums are verified. The initial runtime appended its completion event after hashing its log; the immutable scientific files still match. SHA256SUMS.json hashes the finalized release, including the completed logs.

Reproduce and load

from datasets import load_dataset
measurements = load_dataset("ai-safety-brazil/moderabr-forensics-br", "json_decision", split="analysis", revision="PIN_RELEASE_COMMIT")

Pin the actual release commit. code/ contains both exact executed GPU scripts, the numerical adapter converter, the PEFT meta-device preflight and analysis code. Use the recorded base/adapter revisions and converted-weight checksums. The execution adapter is a compatibility conversion of the original rank-32 LoRA: separate Q/K/V A matrices are concatenated and B matrices block-diagonalized, with rank/alpha tripled for the fused module; the update scale is preserved. The output projection alias maps to lm_head. See conversion.json; no SVD, truncation or retraining.

For GPU reproduction, download the verified compatible model export (including meta-preflight.json), set MODERABR_ADAPTER_DIR to that folder, and run the corresponding PEP 723 script with uv run. Set FORENSICS_FOLLOWUP=decision-sites for code/json_decision.py; leave it unset for code/instruction_prefix.py.

Source adapter: ai-safety-brazil/moderabr-4b@2e370683bb7d9ced0645c8d035b18818f11fae1a. Base: Qwen/Qwen3.5-4B@851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. The verified compatible export is published in the ModeraBR model repository; publication receipts identify its exact revision and weight hash.

License and citation

Research data and documentation are CC BY-SA 4.0. The adapter conversion retains the original ModeraBR CC BY-SA 4.0 license. The base model retains its own license; base weights are not redistributed here. Code retains the repository's MIT license. Cite this release and the source model revisions. These citation requests add no restrictions beyond the applicable licenses. Keep the original model/source notices.

Citation and attribution

Please cite this dataset in publications, evaluation reports, model cards and other published work that uses it. State the dataset version and the actual Hugging Face commit SHA used, so results remain reproducible. CITATION.cff and CITATION.bib provide machine-readable citation metadata.

Preserve author attribution and the existing licence notices when redistributing this work. See LICENSE; these citation instructions add no new use restrictions.

@dataset{aisafetybrazil2026_moderabr_forensics_br,
  author = {{AI Safety Brazil}},
  title = {moderabr-forensics-br},
  year = {2026},
  url = {https://huggingface.co/datasets/ai-safety-brazil/moderabr-forensics-br},
  version = {0.1.0},
}