MiniCPM5-1B PT-BR v3 (instruct)

Brazilian Portuguese instruct model built from openbmb/MiniCPM5-1B-Base:

  1. QLoRA continued pretraining on Portuguese Wikipedia
  2. Harness-aligned Open PT SFT (41k examples, assistant-only loss)
  3. LoRA merge to full weights

No WhatsApp or client data.

Open Portuguese LLM Leaderboard (local eval, not submitted)

Task Base v3
ENEM 2.9% 37.7%
BLUEX 4.2% 34.2%
OAB 3.2% 31.8%
ASSIN RTE 68.3% 71.3%
ASSIN STS 40.5% 48.0%
FaQuAD NLI 54.6% 58.9%
HateBR 75.1% 76.1%
Hate Speech 56.0% 56.1%
TweetSentBR 42.6% 41.7%
Average 38.6% 50.7%

Reference: meta-llama/Meta-Llama-3.1-8B-Instruct scores 70.2% on the same suite.

Downloads last month
113
Safetensors
Model size
1B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for marinarosa/MiniCPM5-1B-PTBR-v3

Finetuned
(1)
this model
Finetunes
1 model
Quantizations
1 model