12th of July Updates

[alebjanes]

RAG Evaluation

1. 100 questions
   Types of questions:

• 60 on general trade
• 12 on growth/variation
• 28 on rankings

2. RAG evaluation results

Best combination tested so far: multi-qa-mpnet-base-cos-v1 (embeddings) + gpt-3.5-turbo (LLM)

• Accuracy: 73%
  • Answers missing data: 9
  • Answers missing context: 14
  • Incorrect answers: 4
• Average latency: 4.19 s

Out of the wrong answers:

• 24 were general questions
• 3 of growth (the lowest)
• 0 of ranking questions

We're preparing a presentation gathering the results of all approaches with more detail. Next week I'll be improving the RAG + LLM and evaluating the previous multi-layer approach with Pippo.

[pippo-sci]

Fine-tune MAPE results:

• Recap: Accuracy is 0% for all fine-tuned versions because numbers aren't right
• Measuring the absolute percentage error by model
  

Measuring the median absolute % error on value qty, taking the 1.1B params model with one epoch as baseline (ft_tiny0), the error decreased on the 50 epoch trained model (ft_tiny2) and increased when increasing model size (to 7B params, ft_llama2). Sort of make sense, larger models are harder to finetune.

Fine tuning tinyLlama to produce API calls instead, the accuracy is 12%, many because it struggles with HS code numbers. Other than that, query accuracy is 89%. On the HS numbers the % mean error is about 12%. but again every time the model is query with the same question it returns a slightly different number.