Fine-Tuning

Defend pinning the tokenizer revision alongside the model weights when shipping a fine-tune

Why does completion-only loss usually outperform full-sequence loss on the same SFT data?

Walk through what happens during backprop for a LoRA-wrapped linear layer

Sequence the forward passes and loss steps inside a single DPO training step

Order the pieces of state that must be restored to resume a training run bit-exactly

Tulu-style instruction tuning trades volume for diversity: which axes of diversity matter most?

Walk through the triage order when a customer-support LoRA tanks MMLU by 4 points post-deploy

Pick the most likely cause when training loss drops for 1k steps, plateaus for 2k, then drops again at step 3k

A small medical fine-tune leaks verbatim credit-card snippets under red-teaming: identify the failure mode and the standard mitigation

LoRA freezes the base weights: does that mean activation checkpointing on the base is unnecessary?

Merging a LoRA back into an NF4-quantized base hits one subtle problem: name it and explain why

Adam stores two fp32 buffers per trainable parameter: what does that imply for LoRA vs full FT on a 7B?

Spot the bug: packed-sequence SFT where the model learns to copy across unrelated examples

Why does using the same model family as synthetic-data teacher AND eval judge inflate FT scores?

When eval prompts appear verbatim inside an SFT file after a synthetic-data run, what's the root cause?

A distilled model hits 92% on a public benchmark but 31% on an internal hold-out: diagnose.

Spot the bug: chat-template code that drops the EOS from the labels.

QLoRA OOMs at step 7000 after one clean epoch on a 24 GB GPU: which non-obvious causes deserve a look?

Why might a LoRA that passes every eval check then garble one in five prod requests?

RLOO's name hides its baseline trick: what does 'Leave-One-Out' literally compute?

Qwen 3.5 SFT data: fill the special tokens that wrap an assistant turn.

Estimate the trainable param count for a Llama-3 8B LoRA at r=16 targeting all-linear modules.

With max_grad_norm=1.0 and a raw gradient L2 of 12.5, what is the post-clip norm and scale factor?

Same LoRA recipe on H100 vs A100: what precision should each team default to?

Only 200 examples for a niche format task: which PEFT method gives the best shot?