Skip to content
Back to search

axion1_0

Frontier-level agent capabilities with nothing more than high-quality data and a simple training recipe: 2 SFT + RL, no mid-training/DPO/OPD. Full recipe released.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today

Check it deeper on your machine

It has changed since its scan on 11 Nov 2026.

As a Claude Code plugin, install it and then ask Claude Code:

/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/bytedance/axion1_0 with scanskillsafety

Or from a terminal, clone it and run it:

git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/bytedance/axion1_0'

It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/bytedance--axion1-0/badge.svg)](/listings/bytedance--axion1-0)

HTML

<a href="/listings/bytedance--axion1-0"><img src="/api/listings/bytedance--axion1-0/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Clone the repository and follow the [training guide](axion/training/README.md) for environment setup, two-stage SFT and checkpoint export. SFT uses Swift with Megatron-FSDP; RL uses verl. They implement successive stages of the same training pipeline.

git clone https://github.com/bytedance/axion1_0.git
cd axion1_0
axion/
  training/
    sft/         Two-stage SFT, runtime patches and checkpoint export
    rl/          Agent rollouts, rewards, tools and RL configuration
  datagen/       Stage I and Stage II data construction
  evaluation/    Benchmark preparation and grading
third_party/
  swift/         SFT backend
  verl/          RL backend
tests/           Regression and packaging checks

The 122B pipeline consists of two-stage SFT followed by asynchronous GRPO. The training guides cover single-host and multi-host layouts, required services, checkpoint export, and runtime requirements.

For data construction and evaluation, use a separate Python 3.11+ environment:

python -m pip install -e ".[data,evaluation]"

4.1 Inference

Export a native SFT checkpoint to Hugging Face format before serving it. The [inference guide](axion/evaluation/README.md#serve-a-checkpoint) provides a vLLM launch command and an OpenAI-compatible request example. Agent evaluation also requires a tool-execution harness.

4.2 Documentation

| Guide | Contents | |:---|:---| | [Stage I data](axion/datagen/stage1/README.md) | Problem generation, rewriting, answer verification, and trajectory filtering | | [Stage II data](axion/datagen/stage2/README.md) | Database generators, question filtering, scoring, and trajectory processing | | [SFT](axion/training/sft/README.md) | Runtime setup, data requirements, two-stage training, and checkpoint export | | [RL](axion/training/rl/README.md) | Paper recipe, task selection, services, and launch | | [Evaluation](axion/evaluation/README.md) | Dataset preparation and grading for FrontierScience, HiPhO, HLE-130, and SEAL-0 |

Evaluation scripts prepare datasets and grade rollouts collected with an external agent harness. The evaluation guide documents the required format.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 210 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. <div align="center">
  2. <picture>
  3. <source media="(prefers-color-scheme: dark)" srcset="assets/axion-logo-dark.svg">
  4. <img src="assets/axion-logo.svg" width="260" alt="Axion 1.0">
  5. </picture>
  6. <h1>Open Agent Foundation Models, Simplified</h1>
  7. <p>
  8. <a href="https://axion1-0.github.io/"><img src="https://img.shields.io/badge/Project-Website-2563EB?style=flat-square" alt="Project website"></a>
  9. <a href="https://axion1-0.github.io/paper/axion.pdf"><img src="https://img.shields.io/badge/Read-Technical_Report-0F766E?style=flat-square" alt="Technical report"></a>
  10. <a href="https://huggingface.co/datasets/Axion1-0/axion-training-data"><img src="https://img.shields.io/badge/Data-Hugging_Face-0F766E?style=flat-square" alt="Training data on Hugging Face"></a>
  11. <a href="https://huggingface.co/Axion1-0/Axion-1.0-35B"><img src="https://img.shields.io/badge/Model-35B_Reproduction-2563EB?style=flat-square" alt="Axion 1.0-35B reproduction model on Hugging Face"></a>
  12. <a href="LICENSE"><img src="https://img.shields.io/badge/License-Apache_2.0-2563EB?style=flat-square" alt="Apache 2.0 license"></a>
  13. </p>
  14. <p>
  15. <a href="https://jyqhahah.github.io/">Yuqi Jia</a><sup>1</sup> · <a href="https://sunsmarterjie.github.io/">Yunjie Tian</a><sup>2,†</sup> · <a href="https://www.linkedin.com/in/jiayu-zheng-5aa231250/">Jiayu Zheng</a><sup>2</sup> · <a href="https://joliang17.github.io/">Yijun Liang</a><sup>3</sup> · <a href="https://jingchensun.github.io/">Jingchen Sun</a><sup>4</sup> · <a href="https://www.linkedin.com/in/mengyu-zhao-7ab19274/">Mengyu Zhao</a><sup>2</sup> · Di Fu<sup>2</sup>
  16. </p>
  17. <p>
  18. <sup>1</sup> Duke University &nbsp; <sup>2</sup> ByteDance &nbsp; <sup>3</sup> University of Maryland, College Park &nbsp; <sup>4</sup> University at Buffalo<br>
  19. <sup>†</sup> Project Lead
  20. </p>
  21. </div>
  22. ## 1. Highlights
  23. - **Three-stage training.** Two stages of supervised fine-tuning (SFT) followed
  24. by reinforcement learning (RL), without dedicated mid-training, DPO, or
  25. on-policy distillation.
  26. - **Scientific reasoning and tool use.** Axion 1.0-122B scores **59.0** on full
  27. HLE with restricted website access, **87.0** on FrontierScience Olympiad, and
  28. **66.7** on FrontierScience Research. Axion 1.0-35B scores **62.2** on SEAL-0.
  29. - **Verifiable training data.** Reasoning problems are strengthened using
  30. feedback from the base model. Tasks requiring tools are constructed from
  31. public data sources with programmatically determined reference answers.
  32. - **Open training code.** The repository includes data generation and
  33. filtering, prompts, two-stage SFT, asynchronous RL, and benchmark preparation
  34. and grading.
  35. ## 2. Results
  36. <p align="center">
  37. <a href="assets/benchmark-results.png">
  38. <img src="assets/benchmark-results.png" width="100%" alt="Axion 1.0 compared with selected open and proprietary models on HLE, BioMysteryBench, SEAL-0, HiPhO, and both FrontierScience splits">
  39. </a>
  40. </p>
  41. Selected results from Table 1 of the
  42. [technical report](https://axion1-0.github.io/paper/axion.pdf#page=9):
  43. | Model | HLE | BMB | SEAL-0 | HiPhO | FS Olympiad | FS Research |
  44. |:---|---:|---:|---:|---:|---:|---:|
  45. | Agents-A1 | 47.6 | — | 56.4 | 46.4 | 79.0 | 40.0 |
  46. | Apodex 1.1 (Agent Team) | 56.1 | 35.3 | — | — | — | 63.3 |
  47. | GPT-5.6 Sol | — | 47.1* | 53.2* | 42.9* | 79.0* | 63.3* |
  48. | Gemini 3.1 Pro | 51.4 | 35.3* | 53.2* | 38.5* | 85.0* | 26.7* |
  49. | Claude Fable 5 | 63.8 | 46.1 | — | — | — | — |
  50. | **Axion 1.0-35B** | **52.2** | **35.3** | **62.2** | **45.7** | **86.0** | **56.7** |
  51. | **Axion 1.0-122B** | **59.0** | **41.2** | **58.6** | **46.2** | **87.0** | **66.7** |
  52. BMB denotes the human-difficult subset of BioMysteryBench; FS denotes
  53. FrontierScience. * marks baselines evaluated by the Axion team; other baseline
  54. scores are from official reports. — denotes an unreported result.
  55. ### 2.1 Models
  56. | Model | Base model | Total / active parameters |
  57. |:---|:---|:---:|
  58. | Axion 1.0-35B | Qwen3.6-35B-A3B | 35B / 3B |
  59. | Axion 1.0-122B | Qwen3.5-122B-A10B | 122B / 10B |
  60. Both models in the paper use SFT and RL. The paper also reports 4B and 9B
  61. variants trained with SFT only.
  62. We encourage users to try the
  63. [Axion 1.0-35B community reproduction](https://huggingface.co/Axion1-0/Axion-1.0-35B),
  64. announced in [issue #1](https://github.com/bytedance/axion1_0/issues/1).
  65. This is a community release, not an official Axion checkpoint.
  66. See the model card for reproduction results and inference instructions.
  67. Training data is available on
  68. [Hugging Face](https://huggingface.co/datasets/Axion1-0/axion-training-data).
  69. ## 3. Training Recipe
  70. <p align="center">
  71. <a href="assets/data-construction.svg">
  72. <img src="assets/data-construction.svg" width="100%" alt="Two-stage data construction: model-aware strengthening of self-contained reasoning problems, followed by deterministic construction and verification of tool-dependent tasks">
  73. </a>
  74. </p>
  75. ### 3.1 Stage I: Self-Contained Reasoning
  76. Stage I uses **10,819 trajectories** of scientific and mathematical reasoning.
  77. Public seed problems are rewritten into harder questions. Problems solved by
  78. the base model are further strengthened through composition, parameterization,
  79. range shifts, or case analysis, with reference answers checked after each
  80. rewrite. Solvers may use Python, but not search or webpage retrieval.
  81. ### 3.2 Stage II: Tool-Dependent Reasoning
  82. Stage II uses **7,544 trajectories**, approximately **312M tokens**. Programs
  83. filter records, select a unique entity, and compute a reference answer before
  84. an LLM formulates the question. Solvers must find the required evidence through
  85. tools; construction data and answers are withheld during solving. Questions
  86. solvable without retrieval are removed.
  87. Generators cover individual entities, ranked lists, and chains across
  88. databases, including the NASA Exoplanet Archive, OurAirports, OpenAlex,
  89. UniProt, and Ensembl.
  90. ### 3.3 Reinforcement Learning
  91. Starting from the Stage II checkpoint, **asynchronous GRPO** trains on agent
  92. interactions with search, webpage visits, scholarly retrieval, and Python.
  93. Correct final answers receive full reward; verified intermediate progress can
  94. receive partial credit. Context condensation divides long rollouts into
  95. sessions for training.
  96. ### 3.4 Training Infrastructure
  97. The paper's SFT runs use a 128K context window with FSDP, expert parallelism,
  98. and context parallelism. RL separates rollout workers, trainers, and auxiliary
  99. services into distinct GPU pools. Token-In-Token-Out (TITO) preserves token
  100. IDs, and Rollout Routing Replay (R3) aligns MoE expert choices between rollout
  101. and training.
  102. Two-stage SFT costs reported in Section 5.1:
  103. | Model | GPUs | Stage I | Stage II | Total GPU-hours |
  104. |:---|---:|---:|---:|---:|
  105. | Axion 1.0-35B | 32 | 1h 02m | 1h 46m | ≈90 |
  106. | Axion 1.0-122B | 32 | 3h 32m | 3h 58m | 240 |
  107. These costs are reported in the paper. The released 122B SFT recipe uses
  108. 32 GPUs, a 128K context window, and a global batch of 8. Its saved training
  109. arguments record micro-batch size 1; Section 5.1 lists 2. The released
  110. [SFT recipe](axion/training/sft/README.md) preserves the saved training arguments.
  111. ## 4. Getting Started
  112. Clone the repository and follow the [training guide](axion/training/README.md)
  113. for environment setup, two-stage SFT and checkpoint export. SFT uses Swift with
  114. Megatron-FSDP; RL uses verl. They implement successive stages of the same
  115. training pipeline.
  116. ```bash
  117. git clone https://github.com/bytedance/axion1_0.git
  118. cd axion1_0
  119. ```
  120. ```text
  121. axion/
  122. training/
  123. sft/ Two-stage SFT, runtime patches and checkpoint export
  124. rl/ Agent rollouts, rewards, tools and RL configuration
  125. datagen/ Stage I and Stage II data construction
  126. evaluation/ Benchmark preparation and grading
  127. third_party/
  128. swift/ SFT backend
  129. verl/ RL backend
  130. tests/ Regression and packaging checks
  131. ```
  132. The 122B pipeline consists of two-stage SFT followed by asynchronous GRPO.
  133. The training guides cover single-host and multi-host layouts, required services,
  134. checkpoint export, and runtime requirements.
  135. For data construction and evaluation, use a separate Python 3.11+ environment:
  136. ```bash
  137. python -m pip install -e ".[data,evaluation]"
  138. ```
  139. ### 4.1 Inference
  140. Export a native SFT checkpoint to Hugging Face format before serving it. The
  141. [inference guide](axion/evaluation/README.md#serve-a-checkpoint) provides a vLLM
  142. launch command and an OpenAI-compatible request example. Agent evaluation
  143. also requires a tool-execution harness.
  144. ### 4.2 Documentation
  145. | Guide | Contents |
  146. |:---|:---|
  147. | [Stage I data](axion/datagen/stage1/README.md) | Problem generation, rewriting, answer verification, and trajectory filtering |
  148. | [Stage II data](axion/datagen/stage2/README.md) | Database generators, question filtering, scoring, and trajectory processing |
  149. | [SFT](axion/training/sft/README.md) | Runtime setup, data requirements, two-stage training, and checkpoint export |
  150. | [RL](axion/training/rl/README.md) | Paper recipe, task selection, services, and launch |
  151. | [Evaluation](axion/evaluation/README.md) | Dataset preparation and grading for FrontierScience, HiPhO, HLE-130, and SEAL-0 |
  152. Evaluation scripts prepare datasets and grade rollouts collected with an
  153. external agent harness. The evaluation guide documents the required format.
  154. ## 5. License
  155. [Apache 2.0](LICENSE).
  156. ## 6. Citation
  157. ```bibtex
  158. @article{jia2026axion,
  159. title={Axion 1.0: Open Agent Foundation Models, Simplified},
  160. author={Jia, Yuqi and Tian, Yunjie and Zheng, Jiayu and Liang, Yijun and Sun, Jingchen and Zhao, Mengyu and Fu, Di},
  161. journal={arXiv preprint arXiv:2611.03233},
  162. year={2026}
  163. }
  164. ```

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.