ner-pt-f1-v1-qwen35-2b-specific-paramopama-seed3407

LoRA adapter for generative named-entity recognition in Portuguese. This artifact is one run of the ner-pt-generative-2026-f1-v1 research matrix.

Run configuration

• Base model: Qwen/Qwen3.5-2B@15852e8c16360a2fea060d615a32b45270f8a8fc
• Regime: specific
• Dataset: paramopama
• Seed: 3407
• Precision: BF16 training with LoRA
• Checkpoint selection: validation end-to-end F1; the test split was not used for selection
• Canonical inference: vLLM, temperature 0, labels_and_tokens constrained JSON
• Invalid-output policy: empty prediction in end-to-end scoring

Test results

Dataset	Precision	Recall	F1	Structural validity
paramopama	0.8963	0.8978	0.8971	0.9988

Results describe only the listed frozen splits and this single seed. Cross-seed uncertainty requires completion of the full three-seed matrix. Results must not be interpreted as evidence of general performance outside these corpora.

Artifacts and reproducibility

The repository root contains the LoRA adapter. research/ contains frozen predictions, normal and decontaminated metrics, schema ablations, inference contracts, the run manifest, hashes, and checkpoint-selection evidence.

Load this adapter with PEFT on the exact base-model revision above. Reproduce the documented protocol with BF16-capable hardware and the versions recorded in research/manifest.json.

Intended use and limitations

Intended for Portuguese NER research, evaluation, and controlled downstream experimentation. Generated spans can be structurally valid yet semantically wrong. Corpus annotation schemes differ, text overlap may affect estimates, and this model has not been validated for high-stakes or autonomous decisions. Users must review dataset licenses and evaluate errors in their target domain.