Design A/B rollout analysis and drift detection for citation accuracy on a production LLM app in code assistant.
Full fine-tuning recipe: QLoRA (4-bit) on DeepSeek-V3 base via Unsloth, targeting single A100 80GB, with data mix and eval plan.
Full fine-tuning recipe: QLoRA (4-bit) on DeepSeek-V3 base via LitGPT, targeting 8x H100, with data mix and eval plan.
Full fine-tuning recipe: QLoRA (4-bit) on Gemma 2 27B via OpenRLHF, targeting AWS g5.12xlarge, with data mix and eval plan.
Full fine-tuning recipe: QLoRA (4-bit) on Gemma 2 27B via torchtune, targeting 2x A100 80GB, with data mix and eval plan.
Full fine-tuning recipe: QLoRA (4-bit) on Gemma 2 27B via Megatron-LM, targeting single RTX 3090 (24GB), with data mix and eval plan.
Full fine-tuning recipe: QLoRA (4-bit) on Qwen 2.5 32B via Unsloth, targeting Lambda Labs 8xH100, with data mix and eval plan.
Full fine-tuning recipe: DPO on Gemma 2 27B via torchtune, targeting Lambda Labs 8xH100, with data mix and eval plan.
Full fine-tuning recipe: DPO on Phi-4 via Unsloth, targeting single A100 80GB, with data mix and eval plan.
Full fine-tuning recipe: DPO on DeepSeek-V3 base via OpenRLHF, targeting single A100 80GB, with data mix and eval plan.
Full fine-tuning recipe: DPO on Mixtral 8x22B via DeepSpeed, targeting single H100 80GB, with data mix and eval plan.
Full fine-tuning recipe: DPO on Yi 1.5 34B via Hugging Face TRL, targeting single H100 80GB, with data mix and eval plan.