← All methodologies

Supervised Fine-Tuning

SFT
Start by learning from worked examples.

Give the model examples of the answers you want. Train it to predict those answers. This is supervised learning, not reinforcement learning.

The idea

Before solving problems alone, you study a teacher's worked solutions. You learn what a good solution looks like by following examples rather than by trying many answers and earning scores.

A target answer teaches the model what to write.

Worked example · SFT6 steps

Step 1 of 6

Start with a worked example.

SFT data is a list of conversations written or approved by people. Each one pairs a prompt with the exact answer we want the model to learn to give.

Step 2 of 6

Cut the text into tokens.

The model reads and writes tokens: words or pieces of words. Only the assistant’s tokens are graded. The question’s tokens are context the model reads but is never asked to produce.

Step 3 of 6

At each spot, the model guesses the next token.

Given “Six groups of”, the model must predict what comes next. The example says “ seven”, but the model gives that only a 30% chance and spreads the rest over “ six”, “ eight” and everything else.

Step 4 of 6

Measure the miss with the loss.

The loss is −log(chance of the right token). A 30% chance gives a loss of 1.20. A 100% chance would give 0. Training is the search for model settings that make the loss small.

Step 5 of 6

Nudge the model toward the example.

Gradient descent adjusts the model’s settings so “ seven” becomes more likely after “Six groups of”. After some training it might rise from 30% to 62%, and the loss falls to 0.48. The same happens at every graded token in every example.

Step 6 of 6

Check it on a question it never saw.

After many examples, the model follows the pattern on new prompts. There was no reward, no judge and no practice attempts: the model copied examples. That is why SFT is supervised learning, not reinforcement learning.

Training example written by a person

userExplain 6 × 7.

assistantSix groups of seven make 42.

Tokens

Explain␣6␣×␣7.Six␣groups␣of␣seven␣make␣42.

question: context, not gradedanswer: gradedthe token to predict␣ marks a space

Model · after “Six groups of” comes…
“ seven” · target62%
“ six”15%
“ eight”8%
anything else15%
Loss = −log(chance of “ seven”)
0.48 −log(0.30) = 1.20 · far from 0−log(0.62) = 0.48 · closer to 0
Test on a new question

userExplain 3 × 4.

modelThree groups of four make 12.

Task or dataModel being trainedGood / chosenBad / rejectedFrozen or hidden

jsonsft.jsonl · one rownew at this step

{"messages": [
  {"role": "user", "content": "Explain 6 × 7."},
  {"role": "assistant", "content": "Six groups of seven make 42."}
]}

jsontokenized examplenew at this step

{
  "tokens": ["Explain", " 6", " ×", " 7", ".", "Six", " groups", " of", " seven", " make", " 42", "."],
  "graded": [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
}

In PyTorch the ungraded positions get the label -100, which the loss ignores. Real chat templates also add special tokens that mark who is speaking.

jsonone predictionnew at this step

{
  "context": "Explain 6 × 7. Six groups of",
  "target": " seven",
  "p": {" seven": 0.30, " six": 0.25, " eight": 0.15, "other": 0.30}
}

pythonloss.pynew at this step

import math
loss = -math.log(0.30)   # 1.20

This is called cross-entropy loss. A confident wrong guess costs a lot; a confident right guess costs almost nothing.

pythonloss.pynew at this step

loss_before = -math.log(0.30)   # 1.20
loss_after = -math.log(0.62)    # 0.48

jsonevaluation · new promptnew at this step

{"role": "user", "content": "Explain 3 × 4."}
{"role": "assistant", "content": "Three groups of four make 12."}
One illustrative training step on one token. Real SFT repeats this for every graded token in thousands of examples, in batches. There is no reward and no judge scoring freely generated attempts in this basic loop.

How SFT works

  1. Gather demonstrations

    Collect prompts paired with good target responses.

  2. Predict the next token

    A token is a small piece of text, often part of a word.

  3. Learn from prediction errors

    Adjust the model to better predict the demonstrated response.

  4. Test new prompts

    Check whether it can apply the behavior beyond the examples.

Useful for

Teaching response formats, task instructions, or an initial helpful style before later training.

The catch

Bad examples teach bad habits. Copying a training pattern does not guarantee success on an unfamiliar problem.

A little more detail

Why is SFT in this collection?

Many post-training recipes combine methods. SFT can provide a starting behavior, and a later preference or RL stage can improve it. There is no rule that every model must follow the same sequence.

Who works with SFT

RL environment builders and AI benchmarks with public evidence that they use SFT: their own website says so, or two independent sites do. Only confirmed evidence is shown; each page lists the quotes.

RL environments 18

  • AfterQuery
    Evidence on its own site
    “Our training pipeline had two sequential stages. The first stage teaches the model what good terminal-agent behavior looks like via SFT on gold trajectories. The second stage uses RL to push the model toward solving harder problems it only partially gets right after SFT.”
    afterquery.com
  • Bespoke Labs
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • Collinear
    Evidence on its own site
    “…ods, by relying only on examples from the training data, our model significantly improves both its rationale quality and its scoring accuracy, outperforming similar and even bigger sized models as well as different post-SFT methods, on diverse evaluation benchmarks. Additionally, we also show the quality improvement in rationales through our proposed recipe through human evaluation.”
    blog.collinear.ai
  • Contra Labs
    Evidence on its own site
    “Seed datasets (image, video, UI, motion). Corrective SFT data and RLHF preference pairs.”
    contralabs.com
  • Datacurve
    Evidence on its own site
    “…\\ Get in touch - Agent trajectories \\ \\ Full traces of expert execution, including tool calls, checks, pivots, and recoveries, so agents learn the shape of real work, not just the finished answer.\\ \\ Get in touch - Supervised fine-tuning \\ \\ Demonstrations that set the right behavioral prior for models, captured through bespoke tooling that lets experts work naturally while preserving the operating judgment behind the work.\\ \\ Get in touch”
    datacurve.ai
  • Halluminate
    Evidence on its own site
    “Training has two phases: SFT followed by RL, specifically Group Relative Policy Optimization (GRPO).”
    halluminate.ai
  • Parsewave
    Evidence on its own site
    “The most effective approach is to use curated, real-world engineering tasks that represent a reliable ground truth. Parsewave datasets are designed for use in supervised fine-tuning, reinforcement learning from human feedback, or as benchmarks to assess model performance.”
    parsewave.ai
  • pre.dev
    Evidence on its own site
    “- If reward varies across attempts, train it with RL. - If reward is flat but a strong demonstration scores higher, use targeted SFT to pull the policy off the floor, then probe it again. - If reward is flat and there is no demonstrated margin, drop it.”
    pre.dev
  • Prime Intellect
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • Refresh
    Evidence on its own site
    “…k, and as training gyms with verifiable rewards that models learn from. A verifiable reward is an objective signal of whether the agent completed a task correctly, and it powers the two main ways to post-train an agent: supervised fine-tuning (SFT) on high-quality example trajectories, and reinforcement learning, where the agent improves by acting in the environment and being scored on the outcome. That signal can come from a deterministic verifier, such as…”
    refresh.dev
  • Revelo
    Evidence on its own site
    “Each RL environment ships as a container with the task, a verifier, proof that it can be solved, and its pass rate. SFT data includes chain-of-thought, diffs, and trajectories where you need them.”
    revelo.com
  • Scale AI
    Evidence on its own site
    “…d the BIRD dataset, we demonstrate that RLHF with a hybrid reward model—which combines database execution and LLM evaluation—can improve SQL generation accuracy. Our approach achieves better performance than traditional supervised fine-tuning across most complexity levels by using 75% of training data for fine-tuning and 25% for RLHF with augmentation.”
    scale.com
  • Turing
    Evidence on its own site
    “If you're looking to optimize your LLMs for real-world applications, Turing offers comprehensive LLM training services that integrate SFT, RLHF, and direct preference optimization (DPO). By leveraging high-quality proprietary human data, extensive STEM domain expertise, and advanced training techniques, we’ve helped leading LLM companies and research orga…”
    turing.com
  • Uber AI
    Evidence on its own site
    “Tap into our global network of contributors and specialized experts to generate supervised fine-tuning (SFT) data and execute reinforcement learning and human feedback (RLHF) workflows.”
    uber.com
  • Ulam
    Evidence on its own site
    “Tests shaped verification rewards alone and after specialized SFT, separating proof-critique skill from broad completion reliability.”
    ulam.ai
  • Veris AI
    Evidence on its own site
    “End-to-end SFT on Veris is coming soon. You can export datasets today and fine-tune on external platforms.”
    docs.veris.ai
  • Vmax
    Evidence on its own site
    “On the multi-family holdout, the base Qwen3-8B model solves 11.6% of tasks. GRPO from the base initialization reaches 27.6%. GRPO from the SFT-formatted initialization reaches 43.6%.”
    vmax.ai
  • Wirestock
    Evidence on its own site
    “Explore multimodal dataset samples below, and contact us to access our full library of AI training data spanning pretraining, SFT, RL, and RLHF.”
    wirestock.io

AI benchmarks 82

  • ~1k JavaScript coding tasks in Harbor format
    Evidence on its own site
    “On the multi-family holdout, the base Qwen3-8B model solves 11.6% of tasks. GRPO from the base initialization reaches 27.6%. GRPO from the SFT-formatted initialization reaches 43.6%.”
    vmax.ai
  • Agent-environment papers
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • App-Bench
    Evidence on its own site
    “Our training pipeline had two sequential stages. The first stage teaches the model what good terminal-agent behavior looks like via SFT on gold trajectories. The second stage uses RL to push the model toward solving harder problems it only partially gets right after SFT.”
    afterquery.com
  • AppBench
    Evidence on its own site
    “Our training pipeline had two sequential stages. The first stage teaches the model what good terminal-agent behavior looks like via SFT on gold trajectories. The second stage uses RL to push the model toward solving harder problems it only partially gets right after SFT.”
    afterquery.com
  • Autoresearch
    Evidence on its own site
    “The naive recipe, plain supervised fine-tuning on the lake with the pinned trainer, scores between 0.01 and 0.05. It improves the math but costs the model about 25 points on the regression suite, which puts it on the steep part of the band, where a handful of regres…”
    autoresearch-bench.com
  • AutoResearchExam
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • BankingBench
    Evidence on its own site
    “End-to-end SFT on Veris is coming soon. You can export datasets today and fine-tune on external platforms.”
    docs.veris.ai
  • Bespoke Curator
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • Bespoke-MiniCheck
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • Bespoke-Nimble-9B model
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • BlenderBench
    Evidence on its own site
    “…k, and as training gyms with verifiable rewards that models learn from. A verifiable reward is an objective signal of whether the agent completed a task correctly, and it powers the two main ways to post-train an agent: supervised fine-tuning (SFT) on high-quality example trajectories, and reinforcement learning, where the agent improves by acting in the environment and being scored on the outcome. That signal can come from a deterministic verifier, such as…”
    refresh.dev
  • BrowserBench (2025-11-22):
    Evidence on its own site
    “Training has two phases: SFT followed by RL, specifically Group Relative Policy Optimization (GRPO).”
    halluminate.ai
  • Cats Confuse Reasoning LLM
    Evidence on its own site
    “…12B-23B to match the performance of models in the range of 32B-70B. This blog presents results on using Collinear math correctness curators used by ServiceNow for CPT of Mistral-NeMO-12B. We also show similar gains with SFT on a Phi-3.5B-instruct model using their curated data. The curated datasets and the post-trained models are now open-sourced and available on Hugging Face.”
    blog.collinear.ai
  • computer-1 computer-use harness in Harbor
    Evidence on its own site
    “While supervised fine-tuning gets you far, agentic reinforcement learning with verifiable rewards (RLVR) is how you push models to actually solve problems. We've curated Gauntlet into a focused 4,000-example RLVR dataset, delivered in Harbor format…”
    refresh.dev
  • ConjectureBench: 14,865 open math problems
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • CWE-bench
    Evidence on its own site
    “…ods, by relying only on examples from the training data, our model significantly improves both its rationale quality and its scoring accuracy, outperforming similar and even bigger sized models as well as different post-SFT methods, on diverse evaluation benchmarks. Additionally, we also show the quality improvement in rationales through our proposed recipe through human evaluation.”
    blog.collinear.ai
  • CWE-bench v1
    Evidence on its own site
    “…ods, by relying only on examples from the training data, our model significantly improves both its rationale quality and its scoring accuracy, outperforming similar and even bigger sized models as well as different post-SFT methods, on diverse evaluation benchmarks. Additionally, we also show the quality improvement in rationales through our proposed recipe through human evaluation.”
    blog.collinear.ai
  • data-eng-bench
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • DealTrace Bench (2026-05/06):
    Evidence on its own site
    “Training has two phases: SFT followed by RL, specifically Group Relative Policy Optimization (GRPO).”
    halluminate.ai
  • Design Arena — 3D Design
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Agentic Game Development
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Frontend Web Apps
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — HTML Game Development
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — HTML Slides
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Image Generation
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Image to Web App
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Image to Website
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Logo Design
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — UI Components
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Video Generation
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Web Design Overall
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • Design Arena — Website Builders
    Evidence on its own site
    “According to the model’s companion paper, AesCoder-4B was developed by fine-tuning Qwen3-4B-Instruct-2507 in two stages. The team first carried out supervised fine-tuning on their newly built AesCode-358K dataset. For the Python subset, they refined the VisCode-200K dataset by regenerating its code samples with Qwen3-Coder-480B-A35B-Instruct-FP8, restricting them to essential libraries s…”
    notes.designarena.ai
  • DrugDiscoveryBench with Phylo
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • Dynamic Semantic Tags technical report with Intuit
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • Evalchemy
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • FinanceQA
    Evidence on its own site
    “Our training pipeline had two sequential stages. The first stage teaches the model what good terminal-agent behavior looks like via SFT on gold trajectories. The second stage uses RL to push the model toward solving harder problems it only partially gets right after SFT.”
    afterquery.com
  • General Agent synthetic environment
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.train.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • GEPA
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • GPU Glossary
    Evidence on its own site
    “…policies without consuming a large GPU cluster. It is also a realistic target for end-to-end code-reward experiments where each candidate must be compiled or tested in a Sandbox. For narrow domains, a carefully designed SFT cold start followed by GRPO can outperform a much larger untuned base model, but results depend on data and reward quality.”
    modal.com
  • HiL-Bench
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • Human Creativity Benchmark
    Evidence on its own site
    “Seed datasets (image, video, UI, motion). Corrective SFT data and RLHF preference pairs.”
    contralabs.com
  • Humanity's Last Exam
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • IDE-Bench
    Evidence on its own site
    “Our training pipeline had two sequential stages. The first stage teaches the model what good terminal-agent behavior looks like via SFT on gold trajectories. The second stage uses RL to push the model toward solving harder problems it only partially gets right after SFT.”
    afterquery.com
  • INTELLECT-2
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.train.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • INTELLECT-3 Technical Report
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.train.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • Legora BAR
    Evidence on its own site
    “Our training pipeline had two sequential stages. The first stage teaches the model what good terminal-agent behavior looks like via SFT on gold trajectories. The second stage uses RL to push the model toward solving harder problems it only partially gets right after SFT.”
    afterquery.com
  • LifeSciBench
    Evidence on its own site
    “Supervised fine-tuning (SFT) lets you train an OpenAI model with examples for your specific use case. The result is a customized model that more reliably produces your desired style and content.”
    developers.openai.com
  • Market-Bench
    Evidence on its own site
    “Our training pipeline had two sequential stages. The first stage teaches the model what good terminal-agent behavior looks like via SFT on gold trajectories. The second stage uses RL to push the model toward solving harder problems it only partially gets right after SFT.”
    afterquery.com
  • Math RL Tasks
    Evidence on its own site
    “Tests shaped verification rewards alone and after specialized SFT, separating proof-critique skill from broad completion reliability.”
    ulam.ai
  • MCP Atlas
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • Measuring Autonomous AI Research
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.train.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • OpenThoughts
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • OpenThoughts datasets
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • PRBench
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • PROPEL
    Evidence on its own site
    “On the multi-family holdout, the base Qwen3-8B model solves 11.6% of tasks. GRPO from the base initialization reaches 27.6%. GRPO from the SFT-formatted initialization reaches 43.6%.”
    vmax.ai
  • READY or Not enterprise agent deployment benchmarks
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • Remote Labor Index
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • RSI Bench
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • Rubric-based RL papers
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • Scale Labs (launched 2026-03-09
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • Scaling Agentic RL: 365k+ SWE/terminal/search tasks
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.train.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • Self-rationalization improves LLM as a fine-grained judge
    Evidence on its own site
    “…ods, by relying only on examples from the training data, our model significantly improves both its rationale quality and its scoring accuracy, outperforming similar and even bigger sized models as well as different post-SFT methods, on diverse evaluation benchmarks. Additionally, we also show the quality improvement in rationales through our proposed recipe through human evaluation.”
    blog.collinear.ai
  • Simulated-to-Real Transfer for Web Agents (2026-07-27):
    Evidence on its own site
    “Training has two phases: SFT followed by RL, specifically Group Relative Policy Optimization (GRPO).”
    halluminate.ai
  • SpreadsheetBench 2
    Evidence on its own site
    “Our training pipeline had two sequential stages. The first stage teaches the model what good terminal-agent behavior looks like via SFT on gold trajectories. The second stage uses RL to push the model toward solving harder problems it only partially gets right after SFT.”
    afterquery.com
  • SWE-bench Pro
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • SWE-Bench Pro - Private Dataset
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • SWE-Bench Pro v1 (2025-09-19) and v2
    Evidence on its own site
    “Execution accuracy by model and training method (Base / SFT / GRPO) on the T2SQL benchmark. Dashed line marks the Qwen3-Coder-480B-A35B-Instruct baseline.”
    labs.scale.com
  • Systematic Reward Hacking study
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.train.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • TB3 results
    Evidence on its own site
    “If a reference demonstration scores materially above the policy, the task is teachable. We use the demonstration for a targeted SFT update, then run the rollout ladder again on the new adapter. If reward starts to spread, the task graduates into RL.”
    pre.dev
  • Terminal-Bench 2.0
    Evidence on its own site
    “Can an open coding model be post-trained to work better on a specific repository? How do we use SFT and RL to improve performance? To answer these questions, we curate coding tasks specialized for a repository and post-train Inkling using SFT and RL. Here, we focus on the fontTools open-source repository, but our meth…”
    bespokelabs.ai
  • The Valley of Code Reasoning
    Evidence on its own site
    “…ods, by relying only on examples from the training data, our model significantly improves both its rationale quality and its scoring accuracy, outperforming similar and even bigger sized models as well as different post-SFT methods, on diverse evaluation benchmarks. Additionally, we also show the quality improvement in rationales through our proposed recipe through human evaluation.”
    blog.collinear.ai
  • TRAIL
    Evidence on its own site
    “The typical post-training arc is two stages, SFT → policy-RL , and that is the sequence this article follows too.”
    patronus.ai
  • TraitBasis / Impatient Users Confuse AI Agents
    Evidence on its own site
    “…ods, by relying only on examples from the training data, our model significantly improves both its rationale quality and its scoring accuracy, outperforming similar and even bigger sized models as well as different post-SFT methods, on diverse evaluation benchmarks. Additionally, we also show the quality improvement in rationales through our proposed recipe through human evaluation.”
    blog.collinear.ai
  • True Agents Model the World / ECHO
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.train.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • Universal offline sandbox escape disclosure
    Evidence on its own site
    “A training algorithm in prime-rl is configured under [orchestrator.train.algo] , where type names the algorithm ( grpo , opd , sft , …) and the class defaults are its vetted setting. It has two parts:”
    docs.primeintellect.ai
  • unix-ctf
    Evidence on its own site
    “On the multi-family holdout, the base Qwen3-8B model solves 11.6% of tasks. GRPO from the base initialization reaches 27.6%. GRPO from the SFT-formatted initialization reaches 43.6%.”
    vmax.ai
  • VERITAS
    Evidence on its own site
    “…ods, by relying only on examples from the training data, our model significantly improves both its rationale quality and its scoring accuracy, outperforming similar and even bigger sized models as well as different post-SFT methods, on diverse evaluation benchmarks. Additionally, we also show the quality improvement in rationales through our proposed recipe through human evaluation.”
    blog.collinear.ai
  • Web Bench (2025-05-30; ~5
    Evidence on its own site
    “Training has two phases: SFT followed by RL, specifically Group Relative Policy Optimization (GRPO).”
    halluminate.ai
  • WebBench
    Evidence on its own site
    “Training has two phases: SFT followed by RL, specifically Group Relative Policy Optimization (GRPO).”
    halluminate.ai
  • Westworld Finance Diligence Bench (2026-08-11):
    Evidence on its own site
    “Training has two phases: SFT followed by RL, specifically Group Relative Policy Optimization (GRPO).”
    halluminate.ai
  • Westworld v1 simulated-web environments (2025-11-23):
    Evidence on its own site
    “Training has two phases: SFT followed by RL, specifically Group Relative Policy Optimization (GRPO).”
    halluminate.ai
  • YC-Bench
    Evidence on its own site
    “…ods, by relying only on examples from the training data, our model significantly improves both its rationale quality and its scoring accuracy, outperforming similar and even bigger sized models as well as different post-SFT methods, on diverse evaluation benchmarks. Additionally, we also show the quality improvement in rationales through our proposed recipe through human evaluation.”
    blog.collinear.ai

Read the research

These guides simplify the methods. The papers describe the full training recipes. Examples on this page are illustrative, not experimental results.