a couple of the top of my head:

• LLM in the loop with preference optimization


• synthetic data generation


• cross modality "distillation" / dictionary remapping


• constrained decoding

Traditional LLM fine-tuning requires extensive labeled datasets, creating barriers for smaller teams. DeepSeek R1 RL techniques address this by enabling models to fine-tune on smaller, specialized datasets, which are easier for smaller teams to collect. This is especially valuable in domains like math, where outcomes can be automatically verified

... See more

𝗺𝗲𝘁𝗵𝗼𝗱𝘀 𝗼𝗳 𝗳𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 𝗮𝗻 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲 𝗟𝗟𝗠 𝗲𝘅𝗶𝘀t ↓

- 𝘊𝘰𝘯𝘵𝘪𝘯𝘶𝘦𝘥 𝘱𝘳𝘦-𝘵𝘳𝘢𝘪𝘯𝘪𝘯𝘨: utilize domain-specific data to apply the same pre-training process (next token prediction) on the pre-trained (base) model

- 𝘐𝘯𝘴𝘵𝘳𝘶𝘤𝘵𝘪𝘰𝘯 𝘧𝘪𝘯𝘦-𝘵𝘶𝘯𝘪𝘯𝘨: the pre-trained (base) model is fine-tuned on a Q&A dataset to learn to answer questions

- 𝘚𝘪𝘯𝘨𝘭𝘦-𝘵𝘢𝘴𝘬 𝘧𝘪𝘯𝘦-𝘵𝘶𝘯𝘪𝘯𝘨: the... See more

• Query the RAG anyway and let the LLM itself chose whether to use the the RAG context or its built in knowledge


• Query the RAG but only provide the result to the LLM if it meets some level of relevancy (ie embedding distance) to the question


• Run the LLM both on it's own and with the RAG response, use a heuristic (or another LLM) to pick the best answer

... See more