How I automate my own job at Hugging Face using agents

Talk transcript

1. 0:00

   [upbeat music] Okay.

2. 0:17

   All right. Hello, everyone. Thanks for coming by. Today, I'll talk about how I automate my own job, uh, at Hugging Face using agents.

3. 0:28

   Um, short introduction. I'm just, uh, Niels from Belgium, the land of beer, fries, and chocolate. I studied at KU Leuven, uh, and I'm a machine learning engineer at Hugging Face for five years now.

4. 0:42

   Uh, today, I'll talk about the community science team at Hugging Face, which is the team I'm part of. Uh, then I'll talk about how I automate large parts of the community science team, and finally, uh, also discuss some other efforts, uh, that we do at Hugging Face.

5. 0:58

   So let's start with the community science team at Hugging Face. So basically, this started when I was sen- uh, I was seeing, like, trending research, uh, pop- passing by on GitHub.

6. 1:09

   And a lot of times when I saw new interesting work, um, the weights were not available on Hugging Face, sadly. Like, researchers use either Google Drive or they use GitHub releases, they use Dropbox, they use Zenodo or other servers to put their, um, artifacts on, and this hurts, uh, discoverability of their work.

7. 1:29

   It's, like, not easily, uh, visible or discoverable. And when I then open a GitHub issue to say, like, "Actually, you could put your weights on Hugging Face for free," most of the time people, uh, reply to me like, "Yeah, migrating the weights, uh, from Google Drive to Hugging Face actually makes perfect sense."

8. 1:46

   So yeah, the community science team can also, uh, be described as the Google Drive to the Hub team.

9. 1:52

   Um, why? Because on Hugging Face, we have these paper pages, uh, and every single paper is sourced from arXiv. And then on the right side, you can basically list the linked artifacts, like the, the linked models or datasets, so people can easily, uh, reproduce your paper or find the models or datasets.

10. 2:11

    So yeah, you can see them on the right side. Um, this improves the discoverability of your work because we have these metadata tags or filters on the hub, so you can easily find, for example, a depth estimation model, an LLM if you're interested.

11. 2:25

    You can find them by language. You can tag them, uh, with the library they are compatible with and so on. So this improves the discoverability of your work. So these are, yeah, the metadata tags that you can add to every single model on Hugging Face or every single, uh, dataset.

12. 2:40

    So yeah, this is, like, the main problem that we saw. Like, uh, lots of people, lots of researchers are, like, using third-party services to publish their work. We have the Hugging Face platform, which is like a centralized place where people can find machine learning artifacts.

13. 2:53

    Uh, it also improves with documentation because you can add a model card or a dataset card. We have tooling so you can easily upload or download stuff from Hugging Face.

14. 3:02

    Uh, and it might also help res- uh, researchers in promoting their work. So it's basically a win-win, uh, both for researchers and then other people using the research.

15. 3:13

    So yeah, these are the typical GitHub issues that I was opening. Uh, I always had, like, the same template. I just asked, "Could you please release these checkpoints on Hugging Face?

16. 3:20

    Could you please release this dataset on Hugging Face?" And then I also opened PRs, pull requests, on Hugging Face to add dataset cards or model cards to improve the documentation of those artifacts.

17. 3:33

    But there's a problem. Uh, it's not really scalable for me to open all these GitHub issues or pull requests because every single day there are, like, hundreds of research papers coming out on arXiv, especially now with the AI boom.

18. 3:45

    Um, yeah, also NeurIPS, for example, a major AI conference, they are seeing a massive amount of papers. So can we automate this? Can we scale the community science team with agents?

19. 3:56

    So that's the second part of my talk. How can we, yeah, scale this, uh, to a massive amount of research papers?

20. 4:05

    So the idea is pretty simple. Uh, we should have an AI agent which can help me do this outreach to all these researchers which publish, uh, models or datasets, uh, as part of their research work, and then, yeah, do the outreach in an automated way.

21. 4:21

    So this is the typical workflow that I was following. So basically, whenever I saw a research paper, I first tried to find the GitHub URL, uh, of that paper, if it's available.

22. 4:31

    Then I, I read the README of that GitHub, uh, file, and then I basically check if there's anything new, uh, interesting to be shared on Hugging Face. Uh, it could be that it's on Hugging Face already.

23. 4:43

    In that case, I check whether the model cards or dataset cards are already properly, uh, present, whether the metadata tags, for example, are there. If, uh, not, then I might open a pull request.

24. 4:54

    Otherwise, if the artifacts are not yet on Hugging Face, I open a GitHub issue. And then finally, I also follow up with the author. So that's kind of the workflow that I had to automate, uh, with agents.

25. 5:05

    And there are several ways to solve this. Uh, you could, uh, go with a workflow. Uh, these pictures are, by the way, taken from the blog post, Building effective agents by Anthropic, which is a really great read, uh, read.

26. 5:17

    Um, so on the left side, you see, yeah, a workflow which is more deterministic. You basically use LLM APIs within steps of a predefined path or pipeline, uh, which is more predictable.

27. 5:28

    It's more deterministic. You have more control over it. Of course, it's less flex- flexible. And then on the other hand, you could have a fully-fledged autonom- uh, autonomous agent, which is an LLM in a loop that calls tools until it's done, which is more flexible but also less, uh, predictable.

28. 5:45

    Uh, at the time, yeah, of course, it doesn't have to be a binary story. Uh, you can have a workflow on one hand, you can have a fully autonomous agent on the other hand, but you could, you could of course also mix and match these type of things, uh, for your use case.

29. 5:57

    In my case, I went for, um, a pretty deterministic workflow. Uh, why? Because at the time that I was building this, this was in 2024, was at the time that Anthropic, uh, wrote their blog post, Building effective agents, and there they actually said, "Try to avoid building agents if you really don't have to.

30. 6:15

    Start simple. Start with a single LLM API."

31. 6:19

    Uh, avoid frameworks. Uh, and actually, I think those were great tips. So at the time, I started building a workflow which basically replicated the workflow that I was doing when I was doing this outreach.

32. 6:30

    So yeah, this is the whole, uh, pipeline. This is created using the Excalidraw MCP server and Cursor. It's pretty nice to create a visualization of your code. Uh, I'm not gonna go into the details, but basically it just replicates, um, the workflow that I was doing when doing the outreach.

33. 6:45

    And I use LLM APIs and then each of the steps without any framework, without any agent framework. So it made it quite, uh, deterministic, and I had a lot of control over, uh, how this goes.

34. 6:59

    Um, in terms of deployment of this, uh, workflow, it's a simple cron job. So a cron is just something that runs regularly. In my case, I run it once every night.

35. 7:08

    So when I'm sleeping, there is this agent, but technically it's just a cron job. It's a Python script with an LLM API, which is gonna read all these hundreds of arXiv papers, uh, and then it might open GitHub issues, or it might open pull requests on Hugging Face.

36. 7:23

    I'm using GitHub Actions for this. Uh, I saw this very nice blog post, Free Cron Jobs with GitHub Actions, and actually it's probably the best entry point if you wanna set up cron jobs, um, because GitHub has a pretty generous tier if you wanna get started with, like, putting simple cron jobs, uh, up there.

37. 7:40

    And yeah, it makes it really easy for me, uh, in the UI to manage all these cron jobs. And so yeah, every night I have, uh, hundreds of, uh, GitHub issues being, uh, created.

38. 7:52

    For the tracing part, um, I'm using Langfuse. Uh, yeah, Langfuse also has a, a booth here. Um, Langfuse is pretty great. Um, I use it mostly for the tracing part, so the observability part, just to see what is the LLM doing, what are the inputs, what are the outputs, what are the prompts, uh, how much does it

39. 8:11

    cost, latency, and so on. Um, so yeah, uh, I definitely recommend it.

40. 8:18

    Um, but yeah, as my agents are opening so many GitHub issues every night, I then end up with a l- a massive amount of unread GitHub notifications because people reply to those GitHub issues.

41. 8:30

    And that's, uh, a lot of work to then reply to all of those issues. It's kind of like going through your, uh, mailbox.

42. 8:37

    So you could wonder, could we also, um, automate the follow-up to those GitHub, uh, issues? Because initially I was still -- the GitHub issue creation was done, uh, by an agent, but I was still the one involved in then doing the follow-up.

43. 8:51

    Uh, now a few months ago, I also automated the, the follow-up to those GitHub issues.

44. 8:57

    Again, you could think, how should you solve this? Should you go for a more deterministic workflow, or can you go for a fully autonomous agent, uh, an LLM in a loop which runs with some tools and skills?

45. 9:08

    Um, well, here I went for kind of a fully autonomous agent. Uh, so it's kind of flexible. It's a bit less predictable, but it works quite well. Um, I went for this because, uh, in November of last year at AI Engineer in New York, there was a pretty nice workshop by Anthropic on the Claude Agents SDK.

46. 9:28

    And there they were actually saying that agents might be better than workflows. So they were kind of contradicting themselves. But they also said that models have become so good that you might actually now start to work with fully autonomous agents rather than a workflow.

47. 9:42

    So this is why I went with this approach, and I actually am using the Claude, uh, Agents SDK for this use case. Uh, there was another pretty nice talk, uh, by Cursor, also at AI Engineer.

48. 9:54

    This was in the European version in London a few months ago. There they talked about how they replaced 12,000 lines of custom code, pretty sophisticated workflow, with a very simple 200 lines of code skill.

49. 10:07

    Uh, actually, it's pretty similar for me. Like, I can rep- uh, replace a lot of custom code, thousands of lines of code, with nowadays just a simple agent with maybe a CLI as a tool, uh, and a skill, and that's it.

50. 10:21

    Um, because the models have become so good.

51. 10:25

    So yeah, in terms of the, uh, architecture, this is a bit what it looks like. Um, so it's actually just the Claude Agents SDK, which is, I would say, a pretty good Python SDK for building an agent.

52. 10:37

    Initially, I was using the Claude models. Uh, but then I, since actually this week, I'm using the GLM 5.2 model via Hugging Face inference providers. So Hugging Face does offer a service, uh, which basically wraps a lot of inference providers like Together.ai, Fireworks, Cerebras, and so on.

53. 10:55

    So you can use a lot of open models, uh, in a unified way. It's OpenAI compatible, uh, or Anthropic compatible. And then, uh, I deploy this on a Modal.

54. 11:05

    Modal is also present here today. Um, and it's mainly using, uh, Bash as a tool, so the terminal to, uh, basically, um, do Hugging Face commands because it's using the Hugging Face CLI quite a bit.

55. 11:18

    So I, I combine it with the Hugging Face CLI skill, which is actually all it needs. And then, um, it might com-comment something on GitHub as a follow-up. And it also actually does the posting on Slack because eventually I also want to see the final results on our Slack channel, uh, from Hugging Face.

56. 11:36

    So yeah, given that there's also a lot of hype on GLM 5.2 recently. For example, Cursor, uh, saw great performance on their Cursor Bench. Post-training Bench is another one, uh, where it actually beats Opus 4.8 and it's cheaper.

57. 11:49

    So yeah, there's no reason not to use GLM 5.2, uh, especially given that I work at Hugging Face now. Um, for the deployment, as I said before, I use Modal.

58. 11:58

    Um, it's pretty great if you wanna deploy agents. Uh, in my case, I'm using the batch processing feature. So they allow you to spin up a massive amount of containers all in parallel.

59. 12:09

    Every single container is basically one agent loop that is processing one GitHub issue. Uh, it's super easy to use, I have to say. Um, and the startups are also pretty fast.

60. 12:21

    So I definitely recommend it if you're building, uh, agents that are like, for example, running in the background, running overnight, for example.

61. 12:29

    Um, and then the way I invoke it, yeah, technically, I could also just, uh, deploy this as a cron job model, for example, has support for this. But typically the follow-up on the GitHub issues, I still do that actually manually by invoking it as a skill.

62. 12:44

    So I created a skill for this in Cursor, uh, I call it Process Unread Model. And then what it's gonna do is it's actually gonna invoke an agent, in this case Composer 2.5, which is like the agent that I'm mostly using in Cursor, which is again gonna invoke all the other agents.

63. 13:00

    So that's, this is kind of the loop that people are talking about. And then finally, it's gonna post, uh, all the results on our Slack channel.

64. 13:09

    Uh, so yeah, and this is actually what it just post. So what it does is it basically just post a huge amount of Hugging Face papers, uh, which are these research papers which people can, uh, make available on the Hugging Face, because every time someone mentions it in a model card or dataset card, we index it on

65. 13:26

    the hub. And then it just posts all the artifacts that people have been uploading based on the outreach that we do via GitHub. Um, so yeah, I do this still, uh, in a manual form.

66. 13:36

    So I just invoke the skill and then after a few minutes, these messages, uh, appear on our Slack channel.

67. 13:44

    Um, yeah, I just included some fun results because to be honest, it's quite fun to see people interacting with the agents. Um, to be honest, I don't disclose that it's an agent.

68. 13:56

    Why? Because I think if people know it's a bot, then they might quickly like close the issue. And to be honest, they post exactly the same stuff as I was doing before manually.

69. 14:05

    So I don't actually see any reason to, to do that. Um, so and then you see replies like this, uh, "Hi, Niels. Thanks a lot for your su-suggestion and the clear guidance."

70. 14:16

    I actually also oftentimes see people using an agent to reply to my agents. Uh, so it's kind of the, that internet nowadays. Um, but people, yeah, make all their artifacts available on Hugging Face.

71. 14:27

    And out of the thousands of issues that are being created on Hugging Face, actually so far, I've only had two, uh, negative comments. One [REDACTED:gender] saying, uh, yeah, "Please close this slop."

72. 14:37

    So he closed the issue and then another one. But most of the people, they just say, "Yeah, actually it makes perfect sense to make my weights or my datasets available on Hugging Face.

73. 14:45

    Like, why didn't I think of this?" Um, so it's kind of a win-win, I would say.

74. 14:51

    Uh, I oftentimes I'll also post fun results on our Slack channel. Like for example, one time someone, a researcher from Apple, uh, sent me a DM like, "I saw you reached out to me."

75. 15:01

    But yeah, technically it's my agent just posting a GitHub issue, uh, regarding publishing a new Apple, uh, the artifacts of an Apple paper on Hugging Face. Or for example, it reaches out to Google DeepMind to, uh, publish, um, mathematics datasets.

76. 15:17

    Um, so a lot of times like I receive emails, the one on the right side, where yeah, they want to publish a 400 gigabytes dataset on Hugging Face, but this was also my agent just, uh, opening GitHub issues.

77. 15:29

    Um, yeah, this is another fun result. So, uh, PaddleOCR, it's like a [REDACTED:origin], uh, company. They migrated all their OCR models to Hugging Face based on outreach by, uh, the agents that create is-issues for me.

78. 15:44

    So, um, yeah, it's pretty nice. Another fun result is like whether when it completes the default template of model cards on Hugging Face. So, uh, Meg Mitchell, who also works at, uh, Hugging Face, she has a famous paper called "Model Cards for Model Reporting," making sure that anyone documents their models, uh, in a proper way.

79. 16:04

    And so we do provide this template, which you can see on the left side in the Git diff. And then, uh, the agent is just completing that template based on the content that it finds based on that paper, like the GitHub README, uh, the PDF itself, and so on.

80. 16:21

    Um, yeah, it's also quite funny to see, for example, in this case that it, uh, included me in, uh, the model card. It said, uh, "Model card author's Niels, part of the Hugging Face Community Science team."

81. 16:32

    I never prompted it this way, but it's pretty fun to see. Or people, uh, replying, "Thank you for helping me fix my mistakes." So those are all done by, uh, the agents.

82. 16:44

    Uh, I think the most popular GitHub issue that was created was this, uh, paper, uh, tiny recursive models, which you might have seen was quite trending, um, both on Hugging Face but also on Twitter.

83. 16:56

    Uh, so yeah, more than 60 people actually up voted that issue so that the model was released on Hugging Face. So this is again, I think the win-win, so it's both a win for the researcher making their research more discoverable on Hugging Face, but it's also, yeah, better for the people then who want to build on top

84. 17:12

    of that research and want to use them.

85. 17:16

    Uh, so yeah, I have hundreds of GitHub issues where I think I can show, uh, nice results, um, where people interact with the agents.

86. 17:25

    You might also wonder, yeah, how to avoid slop because you might think, okay, you have an agent, uh, spamming the whole internet with your GitHub issues, like should you even do this?

87. 17:34

    Again, I already talked about the win-win. Um, but a blog post that I highly recommend if you wanna avoid that your agent is just posting slop is, um, the LLM Evals FAQ, uh, by Hamel Hussein.

88. 17:48

    Uh, I would say he's like the main expert when it comes to LLM evaluation. He also has like a, a paid course, but he also publishes a lot of stuff for free online, including this blog post.

89. 17:59

    So I highly recommend to go through it if you wanna learn more about how to evaluate your agents.

90. 18:06

    So my conclusion would be, um, that open models are actually getting great, especially now with GLM 5.2, you have DeepSeek V4 and so on. So, um, yeah, we, we are able to now replace closed source models by open ones.

91. 18:20

    Uh, for my use case, I would say agents are actually better than, uh, workflows. Uh, they only need a single CLI, which is the Hugging Face CLI. They need a single skill, the Hugging Face CLI skill, and a sandbox, and that's all they need to do their work.

92. 18:32

    And finally, yeah, don't forget about evaluation. Um-

93. 18:38

    Finally, uh, I can also discuss some other efforts that we do as part of the Community Science, uh, team, um, very shortly. Um, so I have a Twitter account that I created.

94. 18:49

    It's called Daily Papers, and it actually uses the exact same workflow as my agents behind the scenes to post interesting research papers on X. It, uh, recently crossed 90,000 followers without any involvement of me.

95. 19:02

    I just deployed this, uh, and it posts interesting research papers and artifacts from Hugging Face every four hours or every time someone, uh, releases something cool on Hugging Face.

96. 19:13

    Um, so yeah. And I have like Gemini, uh, determining the best visual to tweet or to include in the tweet. Like, for example, this recent tweet, uh, where it tweeted out that NVIDIA released an optimized version of GLM 5.2, got more than 2,000 likes, so that's pretty cool, uh, to see.

97. 19:32

    And a final effort that I'm working on right now is a revival of, uh, Papers With Code, which is a website that once existed, then it was acquired by Meta, uh, and then sadly it, uh, died.

98. 19:43

    So I'm t- I'm trying to re- revive it in making, uh, research and state-of-the-art easier accessible. Um, for now it lives at paperswithcode.co.

99. 19:54

    Uh, so yeah, you can find benchmarks over there. Uh, for example, for OCR models, own OCR benches like Popular Benchmark. But I'm also, uh, making it an educational resource so that people can learn about technical terms like mid-training, uh, on-policy distillation, and so on.

100. 20:12

     So yeah, uh, that was it for my talk. I hope, uh, you learned something. Thanks all for your attention. [audience applauding] [outro jingle]