Qwen2.5-1.5B - Financial News Impact Analysis (QLoRA, merged)
Fine-tuned from Qwen/Qwen2.5-1.5B-Instruct with QLoRA (4-bit NF4, LoRA r=16, alpha=32, all linear layers)
on 149 synthetic financial news items (fictional events about real companies, generated by openai/gpt-oss-120b).
Given a headline and snippet, it returns JSON: event_type (11 types), primary_entity, mentioned_entities,
impact_direction, magnitude, rationale. Use the system prompt from the training data
(task2_genai/news_prompts.py, ANALYSIS_SYSTEM) at inference.
Best validation loss 0.8619 at epoch 3. Code and evaluation: https://github.com/indeewara/CDAZZDEV-MLE-IndeewaraJayasuriya (task2_genai).
Not investment advice. Labels come from a teacher model, not market outcomes.