specific method · filed under post-training
Instruction hierarchy training
Supervised training on conflicting messages to teach the model to prioritize system instructions over developer instructions and developer instructions over user instructions.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
We collected examples of these different roles of messages conflicting with each other, and supervised gpt-oss to follow the instructions in the system message over developer messages, and instructions in developer messages over user messages.
usedpost trainingin gpt-oss-120b and gpt-oss-20bOpenAI
Filed alongside
Other methods under post-training :: supervised fine-tuning.
Supervised fine-tuningFine-tuningLight SFT on teacher-distribution dataLoRA fine-tuningRejection samplingSelf-correction cold startSFT bootstrapping with synthetic dataAdversarial fine-tuningAutoregressive fine-tuningCold-started SFT modelEvaluation-based early stoppingFine-tuning with demographically balanced datasetsFrozen-network warmup for new special tokensInstruction-following fine-tuningLightweight speaker fine-tuningPrompt-based cold start for tool useReasoning-data system promptSafety training to an internal specificationSFT checkpoint for RL researchSFT on MiMo-generated task dataSpecialist-model trainingSupervised fine-tuning at 256k sequence lengthToken-budget-based SFT data blending