Akshit Kumar's index.html

William: a tiny poetry model in the browser

2026-06-21T00:00:00+00:00

William is a tiny local language model I trained to write short poems. The model on this page is loaded by the browser and sampled locally, one token at a time. There is no server endpoint behind the button.

loading William...

The title line is editable. William tokenizes it in the browser before generating the poem.

William is a small decoder-only transformer: 6 layers, 384 hidden dimensions, 6 attention heads, and a 256-token context window. I trained it locally with MLX on Apple Silicon.

The training pipeline was two-stage. First, the model learned general poem-shaped text from the biglam/gutenberg-poetry-corpus line corpus after filtering out Project Gutenberg boilerplate, headers, editorial apparatus, prose-like blocks, and non-English fragments. Then I fine-tuned it on title/body poem pairs from suayptalha/Poetry-Foundation-Poems, with extra filtering for rows that were too long or too prose-like for the short context window. I also used prism-ml/Bonsai-8B-mlx-1bit locally as a grading model to help reject low-fitness fine-tuning rows and audit pretraining artifacts.

For this page, the MLX checkpoint was converted to ONNX and dynamically quantized to int8. The page downloads that static model file and runs it with ONNX Runtime Web in your browser; the model asset is around 14 MB.

Interactive Shell Inside a Slurm Job

2025-11-15T00:00:00+00:00

To access the shell inside a job, instead of SSHing into the node (in a case when one has multiple jobs in the same node).

srun --jobid=<jobid> --pty bash

Configuring Pyright with uv

2025-11-15T00:00:00+00:00

To get pyright to look at the correct package installation folder (using pyproject.toml, like uv does), add the following to the pyproject.toml.

[tool.pyright]
venv = ".venv"
venvPath = "."

Custom uv Environments and Caches with direnv

2025-10-24T00:00:00+00:00

When working on turing/ada, it is useful to install python packages on the /scratch directory instead of .venv.

Fortunately, the blessed people at uv have a specific command for this.

# .envrc
export UV_PROJECT_ENVIRONMENT=/scratch/venvs/myproj
export UV_CACHE_DIR=/scratch/uv_cache

Add it to direnv and then run direnv allow to approve.

Test using:

uv add tqdm pip # only to be able to >pip show< on >uv run<
uv run python3 -m pip show tqdm

will return something like

...
Location: /scratch/venvs/bitter_lesson/lib/python3.12/site-packages
...

Now the heavy packages can live on cheap compute node storage instead of the master node.

Establishing Communication

2025-06-06T00:00:00+00:00

Michelangelo, The Creation of Adam (c. 1512)

From Ilya Sutskever, Oriol Vinyals & Quoc V. Le, “Sequence to Sequence Learning with Neural Networks,” Advances in Neural Information Processing Systems 27 (2014). arXiv:1409.3215

Section 3.3 - Reversing the Source Sentences

While the LSTM is capable of solving problems with long term dependencies, we discovered that the LSTM learns much better when the source sentences are reversed (the target sentences are not reversed). By doing so, the LSTM’s test perplexity dropped from 5.8 to 4.7, and the test BLEU scores of its decoded translations increased from 25.9 to 30.6.

While we do not have a complete explanation to this phenomenon, we believe that it is caused by the introduction of many short-term dependencies to the dataset. Normally, when we concatenate a source sentence with a target sentence, each word in the source sentence is far from its corresponding word in the target sentence. As a result, the problem has a large “minimal time lag”. By reversing the words in the source sentence, the average distance between corresponding words in the source and target language is unchanged. However, the first few words in the source language are now very close to the first few words in the target language, so the problem’s minimal time lag is greatly reduced. Thus, backpropagation has an easier time “establishing communication” between the source sentence and the target sentence, which in turn results in substantially improved overall performance.

--index-strategy unsafe-best-match

2025-01-17T22:30:47+00:00

Installing PyTorch 2.5.1, Fairseq2, CUDA 12.1.

uv pip install fairseq2 --pre --extra-index-url https://fair.pkg.atmeta.com/fairseq2/whl/pt2.5.1/cu121 --upgrade --index-strategy unsafe-best-match

Note from uv: “A compatible version may be available on a subsequent index (e.g., https://pypi.org/simple). By default, uv will only consider versions that are published on the first index that contains a given package, to avoid dependency confusion attacks. If all indexes are equally trusted, use --index-strategy unsafe-best-match to consider all versions from all indexes, regardless of the order in which they were defined.”

biaffine dependency parsing

2024-11-20T05:25:47+00:00

\[h_{i}^{head} = FNN^{head}(r_i)\] \[h_{i}^{dep} = FNN^{dep}(r_i)\] \[s_{i,j} = Biaff(h_i^{head}, h_j^{dep})\] \[Biaff(x_1, x_2) = x_1^T U x_2 + W(x_1 \oplus x_2) + b\] \[ArcScore :: Biaff(x_1, x_2) \rightarrow p^{|toks|}\] \[LabelScore :: Biaff(x1, x2), pred \rightarrow p^{|labels|}\]

References

Applying Occam’s Razor to Transformer-Based Dependency Parsing: What Works, What Doesn’t, and What is Really Necessary

Fixing the Python REPL in Emacs

2024-10-10T00:00:00+00:00

The default uv python install uses libedit readline which doesn’t work well with emacs.

Fix

pip install gnureadline
python3 -m override_readline
might have to remove __file__ from override_readline.py, this has been solved already here but not merged yet.
this should make the repl use gnureadline.
final check
```
import readline
print(readline.__doc__)
```
should print 'Importing this module enables command line editing using GNU readline.'.

std::multiset

2024-06-24T00:00:00+00:00

For problems needing logarithmic searching and removing, like concert tickets, std::multiset

multiset<int> tickets;
for {
    auto it = tickets.upper_bound(x)
    if () {
        ..
    }
    else {
        tickets.erase(x)
    }
}

Ada Cluster Reference

2024-04-26T00:00:00+00:00

ncdu for looking at the disk usage.
sinteractive -c 10 -g 2 for getting a node.
Jupyter setup
- Run jupyter lab password followed by jupyter lab --no-browser on ada.
- Tunnel to your system: ssh -L 1025:localhost:8888 -J akshit.kumar@ada.iiit.ac.in akshit.kumar@gnodeXX
- connect to localhost:1025 using emacs/browser.
pyenv and pipenv
- torch install pipenv install torch torchaudio torchvision
- set local python as: pyenv local 3.X
- run pipenv shell to use the installed configuration.

Static Types and Legacy Python

2023-11-17T00:00:00+00:00

Update: pls use pathlib

I was given a folder of python2 files to work with for a task required to be done for my lab (LTRC, IIIT-H).

The files, of course, had issues – they were written some time ago by people no longer associated with the project. The issues weren’t really unsolvable in any way.

Most of them involved file_not_found – probably due to a path issue or some variable lacking a parameter.

TypeError: 'bool' object has no attribute '__getitem__'

Most of these issues simply would not have existed if a language with static typing was used, saving time in the future (at this moment).

Fetching Recent Last.fm Tracks

2023-11-06T00:00:00+00:00

#!/usr/bin/env python3

import http.client
import json

api_key = "thistotallyworks"
username = "marxplank"

conn = http.client.HTTPSConnection("ws.audioscrobbler.com")

url = f"/2.0/?method=user.getrecenttracks&user={username}&api_key={api_key}&format=json"
conn.request("GET", url)

response = conn.getresponse()

if response.status == 200:
    data = json.loads(response.read().decode("utf-8"))
    track = data["recenttracks"]["track"][0]
    artist = track["artist"]["#text"]
    track_name = track["name"]

    result = f"{track_name} by {artist}"
else:
    result = "Failed to retrieve recent tracks."

conn.close()

code = (
    "<website code />"
    f'{result}'
    "<rest of it />"
)
from pathlib import Path

root_folder = Path(__file__).parents[0]

rpath = root_folder / "public_html/index.html"
with open(rpath, "w") as f:
    f.seek(0)
    f.write(code)

Typeface preferences and C-k

2023-08-29T00:00:00+00:00

(set-frame-font "-*-Iosevka-regular-normal-normal-*-12-*-*-*-m-0-iso10646-1")

wisdom: C-k does not copy into the global mark-ring most times.

Exporting Org Notes to Jekyll

2022-12-14T15:16:22+00:00

  (defun generate-file-name (title)
	(interactive "sEnter title:")
	(message
	 (format "/Some/Folder/komikat.github.io/_posts/%s-%s.md"
		   (format-time-string "%Y-%m-%d")
				   (replace-regexp-in-string " " "-" (downcase title))))
  )

  (defun generate-header-post (title)
	(format "---
layout: post
title: \"%s\"
date: %s
categories: post
comments: false
---
" title (format-time-string "%Y-%m-%d %H:%M:%S %z")))

then combine the two after converting the org buffer to markdown using ox-gfm.

(defun generate-post (title)
(interactive "sEnter title:")
  (progn
	(org-gfm-export-as-markdown)
	(write-region (generate-header-post title) nil (generate-file-name title) t)
	(write-region nil nil (generate-file-name title) t)
	)
  )

Doesn’t support images for now.

lisp syntax highlight testing

2022-11-14T00:00:00+00:00

(cons "one" "second") ; '("one" "second")

Useful links

Asahi Linux Setup

2022-08-03T12:36:10+00:00

Why? i3 and less clutter

Display Manager and ZSH

LightDM is a mess.
startx after login – remove xdisplay commands and append exec i3 to .xinitrc. ZSH is the default shell.

Emacs

AUR has a native-comp install which works on ARM64.
find it way faster than native-comp on MAC OSX, not sure why.
TODO: custom emacs init functions for mac/linux

IIIT Git setup

IIIT has blocked the default ports on LAN, ```config Host github.com Hostname ssh.github.com Port 443 User git

Host gitlab.com Hostname altssh.gitlab.com Port 443 User git PreferredAuthentications publickey``` after setting username and email.

Email

WIP

Hello world! (for the 7538th time)

2022-01-13T11:04:47+00:00

Update

This website is now serves to be my public notepad.

I’ve been meaning to blog for some time now, finally went around caring enough to actually set it up and deploy. The blog also coincides with January 1st (we’re 13 days late actually but lets ignore that for a moment), a Happy New Years to you too.

If you’re looking to chat, I’m eternally active on discord – komikat#5603, a few other socials are linked in the footer.