Bootcamp/Euclid (JonasThamane): Week 6 - Frontier Model Price Estimation

[NathiJonas]

Summary

Component	Original (OpenAI)	This notebook (Anthropic)
LLM	GPT-4.1-nano (fine-tuned)	Claude Haiku (few-shot)
Training data	60 examples → JSONL → upload → training job	60 examples → JSONL → injected into system prompt
Inference	Fine-tuned model endpoint	Claude API with few-shot system prompt
Evaluation	Tester class + Plotly charts	Identical Tester class + Plotly charts
Dataset	ed-donner/items_lite (HuggingFace)	Same

Key takeaways

• Few-shot prompting with 60 examples is a strong substitute for fine-tuning on small datasets
• The JSONL preparation pipeline is identical and future-proof — if Anthropic releases a fine-tuning API, the same files can be uploaded directly
• Swap claude-haiku-4-5-20251001 → claude-sonnet-4-6 for higher accuracy at higher cost
• Swap MODEL and N_TRAIN at the top of the notebook to experiment

[NathiJonas]

@ranskills

[ranskills]

Thanks, @NathiJonas

REQUESTED CHANGES

1. Clear the outputs from the notebook

[NathiJonas]

Thanks, @NathiJonas

REQUESTED CHANGES

1. Clear the outputs from the notebook

Fixed the output issue. @ranskills

[ranskills]

Thanks, @NathiJonas

Open the PR