insightlens-qwen2.5-1.5b

InsightLens — Text Classification + Sentiment Analysis (fine-tuned Qwen/Qwen2.5-1.5B-Instruct)

Fine-tuned with QLoRA on 6000 examples combining AG News (topic classification) and Amazon Polarity (sentiment).

Intended use

Produces a structured Category / Sentiment / Analysis response for a given text. Designed to be paired with a RAG/chunking layer (see the InsightLens AI project guide) for documents of arbitrary size. No single training example has both a real category and a real sentiment label - the model learns the format and behavior of producing both together, and generalizes this to mixed real-world text at inference time.

Training

• Base model: Qwen/Qwen2.5-1.5B-Instruct
• Method: QLoRA (4-bit NF4), r=16, alpha=32
• Steps: 300, LR: 2e-4, max seq len: 768
• Data: 3000 AG News + 3000 Amazon Polarity examples

Limitations

• English only. Small model - category/sentiment judgments on text very unlike either training distribution (e.g. highly technical or legal text) may be less reliable.
• Category set is influenced by AG News's four categories (World, Sports, Business, Sci/Tech) plus "Product/Customer Review" - the model can generalize beyond these but accuracy on very different domains isn't guaranteed without further fine-tuning.