Sentiment-Analysis-7B: novidades, hardware necessário, como rodar grátis

Microsoft released a 7-billion-parameter model that labels sentiment in a single token. It improves accuracy over its base model by over ten points.

Sentiment-Analysis-7B is a causal language model that reads a user message and outputs a sentiment label followed by a short explanation. The label appears as the first token, so you can get just the classification by limiting the generation to three new tokens. It works for messages aimed at a service, product, or agent, which the system prompt defines as the "Target."

The model starts from the open source Mistral-7B-Instruct-v0.2 checkpoint. Microsoft fine-tuned it using reinforcement learning, where another large language model acted as a judge. The fine-tuning set is small, roughly three hundred examples, and no fixed sentiment labels were used during training. This approach lets the model learn the labeling task directly from the data.

According to the model card, the evaluation metric is accuracy. The original Mistral-7B-Instruct-v0.2 scores 62.0 percent, while the fine-tuned Sentiment-Analysis-7B reaches 72.8 percent. That is an improvement of more than ten percentage points, showing the benefit of the targeted fine-tuning.

The main strength is the higher accuracy for sentiment classification compared with the base model. Because the label is the first token, you can request just three new tokens and get a rapid, low-cost prediction. The model is limited to the three predefined labels and was trained on a small internal dataset, so its performance may vary on domains far from the training examples.

If you run AI models at home and need a free sentiment classifier, this model fits well. It runs on a RTX 4060 in 4-bit mode, and even a RTX 3060 can generate about 49 tokens per second, more than enough for single-token labels. The MIT license lets you download the weights from Hugging Face and integrate them into your own projects without cost.

O que você precisa para rodá-lo em casa? Com 4 bits, ele ocupa cerca de 7 gigabytes de memória. A menor máquina doméstica que o comporta por inteiro é RTX 4060. Em RTX 3060 12 GB com 32 gigabytes de RAM, ele escreve cerca de 49 tokens por segundo. Em RTX 4090 com 64 gigabytes de RAM, ele escreve cerca de 126 tokens por segundo.

Ainda não há API grátis para ele. Os pesos são abertos: você pode baixá-los do Hugging Face sob a licença MIT. A Arena ainda não o classificou: adicionamos a pontuação assim que os votos chegarem.

Compare Sentiment Analysis 7B with every free model, and check what your own GPU can run, on FreeLLMArena.