cemig-rl-releases/energy-gpt-regulatorio-32b-safe-think-v2

Overview

energy-gpt-regulatorio-32b-safe-think-v2 is a 32 billion parameter language model developed by cemig-rl-releases, specifically fine-tuned for the Brazilian electric sector's regulatory domain. It builds upon cemig-rl-releases/energy-gpt-regulatorio-32b-safe-v2, a Qwen3-4B family model already adapted for CEMIG's energy domain.

Key Capabilities and Training

This model underwent a specialized alignment pipeline involving DPO offline followed by GRPO (Generative Reinforcement Learning from Policy Optimization). The primary objective was to simultaneously enhance:

• Safety: Ensuring secure and harmless responses.
• Reasoning Quality: Improving the quality of reasoning for regulatory and technical questions within the energy sector.

Training data for safety alignment included Portuguese translations of Anthropic HH-RLHF's harmless-base datasets. The GRPO stage and RAG evaluation utilized the cemig-rl-releases/questions-Gemma4-31B-rl-source dataset.

Evaluation

The model's performance was assessed across three benchmarks:

• Knowledge: Evaluated using cemig-ceia/energy-eval for response quality (Knowledge Score via LLM-as-judge) and multiple-choice accuracy (Knowledge Accuracy).
• Safety: Classified responses as safe, partially safe, or unsafe using cemig-rl-releases/Safety-Questions-Energy.
• RAG: Assessed task coverage, relative quality, and hallucination rate in retrieval-augmented generation scenarios using cemig-rl-releases/questions-Gemma4-31B-rl-source.

Limitations

This model is intended for internal use within the regulatory energy domain. Responses may contain inaccuracies and require expert review before production deployment.