Full fine-tuning recipe: KTO on Llama 3.1 8B via Unsloth, targeting 2x RTX 4090, with data mix and eval plan.
Full fine-tuning recipe: KTO on Llama 3.1 70B via OpenRLHF, targeting 2x RTX 4090, with data mix and eval plan.
Full fine-tuning recipe: KTO on Mistral Nemo 12B via torchtune, targeting AWS g5.12xlarge, with data mix and eval plan.
Full fine-tuning recipe: KTO on Qwen 2.5 7B via Unsloth, targeting AWS g5.12xlarge, with data mix and eval plan.
Full fine-tuning recipe: KTO on Qwen 2.5-Coder 7B via OpenRLHF, targeting AWS g5.12xlarge, with data mix and eval plan.
Full fine-tuning recipe: KTO on Gemma 2 9B via DeepSpeed, targeting AWS p4d.24xlarge, with data mix and eval plan.
Full fine-tuning recipe: KTO on Phi-3.5-mini via Hugging Face TRL, targeting AWS p4d.24xlarge, with data mix and eval plan.
Full fine-tuning recipe: KTO on Phi-4 via Megatron-LM, targeting Lambda Labs 8xH100, with data mix and eval plan.
Full fine-tuning recipe: KTO on DeepSeek-V3 base via FSDP, targeting Lambda Labs 8xH100, with data mix and eval plan.
Full fine-tuning recipe: KTO on DeepSeek-V3 base via Unsloth, targeting 4x A100 40GB, with data mix and eval plan.
End-to-end SFT dataset construction: collection, labeling, cleaning, dedup, contamination check for legal clause extraction.
End-to-end SFT dataset construction: collection, labeling, cleaning, dedup, contamination check for legal clause extraction.