Open Models Almanac
Methodspost-trainingsupervised fine-tuning

general family · filed under post-training

Supervised fine-tuning

Supervised fine-tuning (SFT) is a post-training stage that adapts a pretrained model on a large, curated mixture of supervised data.1 It trains the model to imitate target responses, teaching it the chat template and capabilities for downstream tasks.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. Supervised Fine-Tuning (SFT): Large general-purpose mixture to provide good capabilities on downstream tasks and teach the chat template — LFM2 Technical Report
  2. we include only the “chosen” responses from preference pairs that we ultimately want the model to imitate during SFT — LFM2 Technical Report

Also called SFT, Supervised Fine Tuning (SFT), identical SFT.

sources
23
models
11
labs adopt it
8
strongest
core

Documented in

Evidence

27 quotes from 23 sources, one card per model, each opening with its strongest.

How sources treat it
evaluated 1optional 1used 24core 1

LFM3-1.7B

Liquid AIadopted4 sources

Large general-purpose mixture to provide good capabilities on downstream tasks and teach the chat template.

corepost trainingin LFM2Liquid AI
4 more from this model

Supervised fine-tuning (SFT): 1.75T tokens across various tasks, including chat, reasoning, tool calling, and agentic traces.

usedpost trainingin LFM3-1.7BLiquid AI

Supervised fine-tuning (SFT): 1.75T tokens across various tasks, including chat, reasoning, tool calling, and agentic traces.

usedpost trainingin LFM3-1.7BLiquid AI

Post-training from this checkpoint (SFT, mixed RL, multi-domain on-policy distillation, and agentic RL) produces LFM3-1.7B.

usedpost trainingin LFM3-1.7BLiquid AI

including supervised fine-tuning, length-normalized direct preference alignment

usedunclearin LFM2Liquid AI

MiMo-V2.5

Xiaomiadopted4 sources

Post-training incorporates SFT, large-scale agentic RL, and Multi-Teacher On-Policy Distillation (MOPD)

usedpost trainingin MiMo-V2.5Xiaomi
3 more from this model

Post-training incorporates SFT, large-scale agentic RL, and Multi-Teacher On-Policy Distillation (MOPD)

usedpost trainingin MiMo-V2.5Xiaomi

supervised fine-tuning and agentic post-training

usedtraining objectivein MiMo-V2.5Xiaomi

general Supervised Fine-Tuning (SFT)

usedpost trainingin MiMo-V2-FlashXiaomi

Super’s post-training pipeline combines supervised fine-tuning

usedpost trainingin NVIDIA Nemotron 3.5 Super and NVIDIA Nemotron 3.5 LightningNVIDIA
2 more from this model

Developers can use NVIDIA NeMo AutoModel for supervised fine-tuning (SFT) and low-rank adaptation (LoRA) to teach domain-specific behavior.

optionalpost trainingin Nemotron 3.5 Super VLNVIDIA

we ran identical SFT on the base checkpoint and on the mid-trained checkpoint and compared the two.

evaluatedpost trainingin Nemotron 3.5 SuperNVIDIA

NVIDIA-Nemotron-3-Ultra-550B-A55B

NVIDIAadopted3 sources

Post-trained with enhanced pipeline involving Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD) for improved model accuracy.

usedpost trainingin Nemotron 3 UltraNVIDIA
5 more from this model

To train the student model, we start from the pre-trained base model and perform Supervised Fine-Tuning (SFT) in two stages

usedpost trainingin Nemotron 3 UltraNVIDIA

post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation

usedpost trainingin Nemotron 3 UltraNVIDIA

To improve upon SFT model, we conduct a unified RLVR

usedpost trainingin Nemotron 3 UltraNVIDIA

post-trained using SFT, RL, and MOPD

usedpost trainingin Nemotron 3 UltraNVIDIA

Supervised Fine-Tuning (SFT) in two stages

usedpost trainingin Nemotron 3 UltraNVIDIA

The model was further fine-tuned on synthetic code, math, science, tool calling, instruction following, structured outputs, and general knowledge data.

usedpost trainingin Nemotron 3.5 LightningNVIDIA
1 more from this model

we ran identical SFT on the base checkpoint and on the mid-trained checkpoint and compared the two.

evaluatedpost trainingin Nemotron 3.5 SuperNVIDIA

Kimi K3

Moonshot AIadopted1 source

The SFT stage establishes a high-quality cold-start policy for the subsequent RL stage.

usedunclearin Kimi K3Moonshot AI

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

↑