Example Use Case Tutorials

These end-to-end examples serve as tutorials for the usage of RapidFire AI’s API. They use publicly available datasets, models, and code.

Attention

These notebooks live on RapidFire AI main and use the shard API: num_shards, save_strategy="shard", and the current RFOptuna() signature. They will not run unchanged against the 0.16.1 build on PyPI. See installing from main.

For RAG and Context Engineering

We have one use case example each for an all-local model, all-OpenAI, and a hybrid workflow: FiQA RAG Q&A chatbot, SciFact RAG for scientific claim verification, and GSM8K few-shot/context engineering for math reasoning, respectively. This set will expand over time to more examples based on community inputs.

You can also check out RapidFire AI RAG on Google Colab instead of your own machine; it showcases a simplified FiQA RAG Q&A chatbot use case.

• FiQA: RAG for Financial Opinion Q&A Chatbot
  • Task, Dataset, and Prompt
  • Model, RAG Components, and Configuration Knobs
  • External Vector Stores: Pinecone and PGVector
• GSM8K: Context Engineering for Math Reasoning
  • Task, Dataset, and Prompt
  • Model, Few-Shot Selection, and Configuration Knobs
• SciFact: RAG for Scientific Claim Verification
  • Task, Dataset, and Prompt
  • Model, RAG Components, and Configuration Knobs

For Fine-Tuning and Post-Training

We have one use case example for each supported control flow from HuggingFace TRL: SFT, DPO, and GRPO. This set will expand over time to more examples based on community inputs. The SFT use case has multiple additional variations showcased, including a “lite” version with a fast SLM and multiple FSDP-based versions with larger models that are auto-partitioned across GPUs.

You can also check out RapidFire AI on Google Colab instead of your own machine; it showcases a simplified SFT use case.

• SFT for Q&A Chatbot
  • Task, Dataset, and Prompt
  • Model, Adapter, and Trainer Knobs
  • Multi-GPU Model Partitioning with FSDP
• DPO for Alignment
  • Task, Dataset, and Prompt
  • Model, Adapter, and Trainer Knobs
• GRPO for Math Reasoning
  • Task, Dataset, and Prompt
  • Model, Adapter, and Trainer Knobs

Adaptive Search with Optuna

Instead of enumerating configs with grid or random search, you can hand config generation over to Optuna, which prunes underperforming runs in flight at shard boundaries and spends the freed resources on new configs. See RFOptuna() on the Multi-Config Specification page for the pruning semantics and the current limitation on the sampler. There are three examples, two for evals mode and one for fit mode.

• Optuna for RAG on FiQA: View on GitHub

• Optuna for RAG on SciFact: View on GitHub

• Optuna for SFT on customer support Q&A: View on GitHub

Multimodal RAG

This example runs RAG over real-world PDFs, parsing text, table, and image elements, summarizing each into embeddable text, offloading the originals to storage, and indexing the summaries in Pinecone. It uses the MMDocIR evaluation dataset and needs the multimodal dependencies from rapidfireai init --multimodal. Also read the Multimodal RAG and Artifact Storage section of the RAG spec page.

• Multimodal RAG on MMDocIR: View on GitHub