Sft

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

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

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

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

Match each instruction dataset to its on-disk row shape.

Walk through one training example for SFT'ing a single weather-tool call.

Replay-buffer mix for domain SFT: how much general data, and from where?

Designing a code-assistant SFT mix: how should domain and general data balance?

Teacher forcing during SFT: which description nails it?

Should the system message contribute to the SFT loss?

Fill in the SFT JSONL: the three required fields of a typical chat example.

Define supervised fine-tuning (SFT) in one breath.

Roles in a chat training example: pick the three that are standard.

Pretraining vs fine-tuning: two key differences in data and objective.

Pick the sensible `num_train_epochs` for a typical instruction-SFT run on a curated dataset

Which loss function does supervised fine-tuning actually minimize?

Base model vs instruction-tuned model: what does the user actually feel?

Alpaca-format examples carry three fields: name them

DeepSeek-style reasoning distillation: what travels from teacher to student, and what gets cut from the loss?

Contrast SFT and continued pretraining when adapting a base model to medical text

Spot the error in this description of SFT loss masking

Explain SFT's loss and the role of prompt-token masking

Where do SFT and DPO sit in the classical RLHF pipeline?

ORPO vs DPO: which architectural difference enables 'single-stage' training?

Explain LIMA's Superficial Alignment Hypothesis and what it implies for your data budget