MIDREAL

Home / deleting-weights

agentic-learning-ai-lab

deleting-weights

WatchingWatchingNone of the above. Nothing stands out yet. More on the methodology page

agentic-learning-ai-lab/deleting-weights

Public code release for the paper "The Surprising Effectiveness of Deleting Weights in LLM Post-Training"

PythonCategories: Models

Our summary

Official code for Deletion Tuning, a post-training method that learns a binary mask over the gate projection of transformer blocks in an LLM, zeroing a small fraction of pretrained weights. The repository covers training with GRPO, SFT, and on-policy self-distillation, saving masks, applying masks to base models, and evaluation.

Who it is for: Researchers working on LLM post-training and continual learning

What the README claims

  • Zeroes fewer than 0.05% of weights and matches full fine-tuning at scale across three post-training regimes zeroing fewer than 0.05% of its weights, with no other modification, matches full fine-tuning (FFT) at scale Not checked
  • Trained masks are hundreds to thousands of times smaller than the corresponding FFT checkpoint the trained mask is hundreds to thousands of times smaller than the corresponding FFT checkpoint Not checked
  • Forgets less than other methods, with prior-task accuracy staying near base level under reinforcement learning its prior-task accuracy stays near the base level while FFT, LoRA, and random sparse fine-tuning lose up to 5.4 points Not checked
  • Method is described as learning which pretrained weights to delete a post-training method that learns which pretrained weights of an LLM to delete Too broad to verify

Written by AI from the project's README on 2026-11-12. It does not affect the verdict.

Watching: no other rule matched yet.

Gaining 29 stars this week. Last push 11 days ago.

Verdict

TOO EARLY scored 1h ago

There is not enough data on this project yet to call it. The numbers below are shown anyway.

Too close to call Hype Gap between -1 and +41.

Early read: we are missing data on the size of the team behind it and how widely it is used. This is an estimate and will be replaced by a verdict once we have the data. How early reads work.

Hype
56
Reality
18
Hype Gap
+38
Momentum
-
Confidence
60%
  • Maintained moves Reality -17 points.
  • Responsive moves Reality -8 points.

Rule verdict-2.0, scoring score-2.0. How we score. Backtest.

Maintained by

Maintained by 1 primary developer, 0 external contributors in the last 90 days.

  • agentic-learning-ai-lab is a GitHub organization.
  • Official organization: the project homepage matches the website on its GitHub profile.
  • GitHub account created November 2023.
  • Top contributors by commits: Jacklu0831 100%.

Why now 22% of its stars arrived in the last 7 days. No Hacker News posts in the last 7 days.

Stars StarsThe number of GitHub stars on the repository at our latest reading.
131
Gain 7d Gain 7dStars gained since our latest reading from at least 7 days ago.
+29
Pace Velocity (Nx usual pace)This week's star gain divided by the project's usual weekly gain. The usual gain is the average weekly gain over the 4 weeks before (days 8 to 35 ago), counted as at least 10 stars. 2.0x means twice the usual pace. Without a reading from 35 days ago, the ranking assumes 2% of the stars the project had 7 days ago as its usual weekly gain. More on the methodology page
2.9x
HN points HN pointsHacker News points added up across posts that mention the project, published in the last 7 days.
0 7 days
Downloads DownloadsDownloads per day of the project's package, as the registry (npm or PyPI) reports them.
n/a
Contributors Contributors, all timePeople and anonymous contributors who have ever committed to the repository, as GitHub counts them. A small number is context, not a flaw: many solid projects are kept by one or two people.
1
Last release Latest releaseDays since the latest GitHub release was published.
n/a

Gain 24h Gain 24hStars gained since our latest reading from at least 24 hours ago. +2. Forks ForksThe number of GitHub forks of the repository at our latest reading. 12. Open issues Open issuesThe number of open issues on the repository at our latest reading, as GitHub counts them. 2. Days since push Days since pushDays since the last push to the repository. 11 days. Archived ArchivedWhether the owner has archived the repository on GitHub. An archived repository is read-only. no. Hacker News posts in 7 days Hacker News postsThe number of Hacker News posts that mention the project, published in the last 7 days.: 0. Watchers WatchersThe number of GitHub users watching the repository for notifications, at our latest reading. 1.

Facts

License
MIT source, checked 22 Nov 2026
Language
Python source, checked 22 Nov 2026
Homepage
https://agenticlearning.ai/deleting-weights/ source, checked 22 Nov 2026
Releases
0 source, checked 21 Nov 2026
Contributors
1 source, checked 21 Nov 2026
Archived
no source, checked 22 Nov 2026

Nerd stuff

Docker

Languages
Python (detected), Shell (detected), Jupyter Notebook (detected), Jinja (detected)
Deployment
Docker (detected)
License
MIT (detected)

The Receipts

SourceEvidenceDate
Newsletter or feed[AINews] D-Day (Decision Models Day)12 Nov 2026
Newsletter or feed[AINews] Mercor $750m Series D at $20B valuation, $3B run rate11 Nov 2026

What changed recently

GitHub release dates we stored. We do not track license changes yet.

Cite this

deleting-weights is a GitHub project at https://github.com/agentic-learning-ai-lab/deleting-weights written mainly in Python, described there as "Public code release for the paper "The Surprising Effectiveness of Deleting Weights in LLM Post-Training"".

Source: https://midorreal.com/project/agentic-learning-ai-lab-deleting-weights, checked 22 November 2026. The same facts as JSON: /project/agentic-learning-ai-lab-deleting-weights.json.

Rank history

Peak day: +23 stars on 13 Nov 2026, 9 days ago.

Stars

11 Nov 202622 Nov 2026
Stars each day Stars each dayThe highest star count we recorded on each day., last 12 days. Without JavaScript this shows the last 30 days.

Events in this window

Dated items we stored, last 30 days. They sit near the chart, they do not explain it.

Stars gained each day Stars gained each dayThe difference in the daily peak star count between consecutive days that both have a reading., between readings on consecutive days. Last 30 days in the fallback.

Packages

- We checked npm, PyPI, Docker Hub, crates.io and Homebrew and found no package that links to this repository.

Signals

SourceWhat we measuredChecked
GitHub131 stars, 29 gained in 7 days22 Nov 2026
Hacker NewsNo posts stored-
Feed mentions2 mentions in tracked newsletters and blogs12 Nov 2026
npmNo linked package found-
PyPINo linked package found-
Docker HubNo linked package found-
crates.ioNo linked package found-
HomebrewNo linked package found-
deps.devNot tracked yet-
Hugging FaceNot tracked yet-
BlueskyNot tracked yet-
Stack ExchangeNot tracked yet-
YouTubeNot tracked yet-

Attention vs adoption

Attention: above 83% of tracked projects 83%
Star velocity against all tracked projects

Adoption: not enough data to rank
Downloads, dependents and pulls

Not enough comparable data yet to set attention against adoption. This describes the numbers; it is not a verdict.

Compared with Models

MeasureThis projectCategory median
Star velocity2.90.6
Days since releaseno data11.1
Contributors, all time12
Median time to first reply on issues27 hours12 hours
Commits in the last 30 days15

Star velocity: above 79% of Models projects

Contributors, all time: above 0% of Models projects

Commits in the last 30 days: above 32% of Models projects

Compared with the project's first category only. A percentile line appears only where at least 30 projects in the category have the measure. This describes the numbers; it is not a verdict.

Compare with

Hacker News

PostPoints PointsThe points the post has on Hacker News.Comments CommentsThe number of comments on the post on Hacker News.Date
No Hacker News posts mention it in the last 7 days.

Caveats

  • GitHub lists 1 contributor to this repository. Source
  • The repository has no security policy on GitHub. Source
  • The repository has 0 releases on GitHub. Source

Repo health

  • Languages LanguagesShare of the repository's code by bytes, as GitHub's language detection reports it. The five biggest languages are named and the rest are grouped as Other.Python 86%, Shell 13%, Jupyter Notebook 1%
  • License LicenseThe SPDX license id GitHub detects, and a class: Open source (OSI) if the license is on the Open Source Initiative approved list, Source-available if the code is readable but the license limits use, Unclear if GitHub cannot match a known license, No license detected if it finds none. A public repository without a license is not open source.MIT, Open source (OSI)
  • Repository age Repository ageTime since the repository was created on GitHub. This is the repository, not necessarily the project, which may have existed elsewhere before.11 days, created 10 Nov 2026

Development

  • Commits in the last 7 days Commits in the last 7 daysCommits to the repository in the last 7 days, as GitHub counts them.0
  • Commits in the last 30 days Commits in the last 30 daysCommits to the repository in the last 30 days, as GitHub counts them.1
  • Commits in the last 90 days Commits in the last 90 daysCommits to the repository in the last 90 days, as GitHub counts them.1
  • Contributors, all time Contributors, all timePeople and anonymous contributors who have ever committed to the repository, as GitHub counts them. A small number is context, not a flaw: many solid projects are kept by one or two people.1, a small team
  • Releases ReleasesThe number of GitHub releases, and how many days ago the latest one was published.0
  • Issues in the last 30 days Issues in the last 30 daysIssues opened and issues closed in the last 30 days, as GitHub search counts them.3 opened, 1 closed
  • Pull requests merged in the last 30 days Pull requests merged in the last 30 daysPull requests merged in the last 30 days, as GitHub search counts them.0

Maintenance

  • Median time to first reply on issues Median time to first reply on issuesMedian time between an issue being opened and the first comment from someone other than its author, over a sample of the last 30 closed issues. Issues with no reply are left out of the median and counted separately.27 hours
  • Median time to close issues Median time to close issuesMedian time from opening to closing, over a sample of the last 30 closed issues.30 hours
  • Sampled issues with no reply Sampled issues with no replyHow many of the sampled closed issues never got a comment from someone other than their author.0 of 1
  • Stale open issues Stale open issuesOpen issues with no update in the last 90 days, as GitHub search counts them.0
  • Open pull requests Open pull requestsPull requests open now, as GitHub search counts them.0

Hygiene

  • Automated tests or builds (CI) Automated tests or builds (CI)Whether the repository has at least one GitHub Actions workflow file in .github/workflows.No
  • A tests folder A tests folderWhether the repository has a top-level folder named tests, test, __tests__ or spec.No
  • A license A licenseWhether GitHub detects a license file in the repository.Yes
  • Contributing guide Contributing guideWhether the repository has a contributing guide.No
  • Code of conduct Code of conductWhether GitHub finds a code of conduct in the repository.No
  • CODEOWNERS file CODEOWNERS fileWhether the repository has a CODEOWNERS file, which names who reviews changes to which files.No
  • A Dockerfile A DockerfileWhether the repository has a Dockerfile at its top level.No

Security

  • Security policy Security policyWhether the repository has a SECURITY.md or SECURITY.rst file.No
  • Unpatched high or critical advisories Unpatched high or critical advisoriesPublished GitHub security advisories rated high or critical on this repository that list a vulnerable version with no patched version.0

Badge

For maintainers: one measured fact, no verdict. First spotted by Mid or Real at 41 stars on 11 Nov 2026

Markdown

[![Mid or Real](https://midorreal.com/badge/agentic-learning-ai-lab-deleting-weights.svg)](https://midorreal.com/project/agentic-learning-ai-lab-deleting-weights)

HTML

<a href="https://midorreal.com/project/agentic-learning-ai-lab-deleting-weights"><img src="https://midorreal.com/badge/agentic-learning-ai-lab-deleting-weights.svg" alt="Mid or Real"></a>

Verdict badge: Mid or Real: too early It shows the verdict as this page shows it today.

Markdown

[![Mid or Real: too early](https://midorreal.com/badge/agentic-learning-ai-lab-deleting-weights.svg?style=verdict)](https://midorreal.com/project/agentic-learning-ai-lab-deleting-weights)

HTML

<a href="https://midorreal.com/project/agentic-learning-ai-lab-deleting-weights"><img src="https://midorreal.com/badge/agentic-learning-ai-lab-deleting-weights.svg?style=verdict" alt="Mid or Real: too early"></a>

Passport

Copy these facts into a post, a README or a page. Cite as "Source: Mid or Real". Full record: JSON.




First tracked 11 Nov 2026 at 41 stars.

Projects blowing up

How we call a verdict

Is this your project? Claim it

Figures by method 1.0