Diensttagebuch

Day 2646 (31 Mar 2026)

Translating pattern with LLMs

I know enough German to judge but not enough to write grammatically myself, and this is a pattern I don’t want to forget.

Pasted image 20260331155128.png

Pasted image 20260331155157.png

This is SO MUCH BETTER than anything else I’ve used for this! For very short snippets it’s golden

Day 2645 (30 Mar 2026)

LaTeX paper checklist

For the main file, see: 240116-1701 LaTeX best practices

Camera-ready checklist

Enable reviewing mode in Overleaf
Change package options [review/final] etc.
Authors emails, names, order correct?
Acknowledgements present and uncommented?
Anonimity: uncensor names+URIs
Are acknowledgements, references, appendices a) allowed, b) part of page limit?
COMPRESSION: reconsider all hacks if you got more space for the camera-ready version
- \setlength{\itemsep}{0pt} \setlength{\parskip}{0pt} and friends
- \setlength{\belowcaptionskip}{0} \setlength{\abovecaptionskip}{10pt}¹ <- default values
- \looseness-1
- Any paragraphs to un-merge for readability?
Go through pre-flight checklist below
Manually check ALL review-mode changes at the end!

Pre-flight checklist

Ask LLMs to proofread! THEN do the steps below.
Acknowledgements, references, appendices: a) allowed, b) count towards page limit?
- In which order?
Anonymity: acknowledgements, again!
Search for TODOs in the text!
Consistency:
- When to use single/double quotes (quotes), when emphasis (foreign languages?)
- terminology?
  - (including in captions and inside figures (labels, titles!))
- Table headers’ variables and names are easy to forget to update!
Tables
- Did any footnote markers get deleted when pasting new data in tables?
- Adding \bottomrules helps if distance between table and caption too small
Titles
- Title Case where needed (paper title + sections)
- Break the paper title in some pretty way
  - \mbox to avoid breaks, \- within word to mark break words
Citations+references
- before periods and footnotes
- always w/ NBSP
- Double-check each reference, including inside table/figure captions!
- citep=cite, citet if subject, citealp in parentheses
- Don’t forget multiple citations exist [1,3]!
Commands/macros
- Search for \TODO and any custom commands/macros you created
- Check the spacing after any custom macros
Typography bits
- 231206-1501 Hyphens vs dashes vs en-dash em-dash minus etc
- Smart Quotes: Either `x’ and ``x’’ or \enquote{} everywhere
- Large Numbers: Write large numbers as 54{,}000.99
- ~ is \textasciitilde: check any estimate numbers (~50%)!
- % is \%; check all percentages!
- \ is \textbackslash
- underscores, subscripts, superscripts are dangerous
- Avoid footnotes after numbers
- does minted caption use the same font as the rest?
- -,--,--- are the safest — look and remove pasted UTF8 ones
NBSP (search and manually check each!)
- Citations, \references (Figure~\ref{fig:somefig})
- TODO
Look at logs!

Things that went wrong content-wise in the past

Read through all analysis and tables, make sure the data makes sense, the analysis/results match the numbers in the tables, the plots and the tables match to each other.
Make sure whatever you used to generate plots and tables used the exact same data!

https://tex.stackexchange.com/questions/23313/how-can-i-reduce-padding-after-figure ↩︎

Day 2643 (27 Mar 2026)

Tiny PRF demo program

zc
zc/it

I wrote this debugging some metrics, bad code but I don’t want it to die. Shows formulas for Precision/Recall/F-Score based on number of TP TN FP FN.

Depends only on rich, runnable w/ uv through 250128-2149 Using uv as shebang line.

Usage: uv run prftset.py tp tn fp fn

Pasted image 20260327031056.png

# /// script
# dependencies = [
#   "rich",
#   "pdbpp", # ви цього варті
# ]
# ///

from rich import print
from rich.layout import Layout
from rich.table import Table
import argparse


def get_scores(tp, tn, fp, fn) -> dict[str, int | float | str | None]:
    """Calculate precision, recall, and F1 score from true positives, true negatives, false positives, and false negatives."""

    precision = tp / (tp + fp) if (tp + fp) > 0 else None
    recall = tp / (tp + fn) if (tp + fn) > 0 else None
    f1_score = (
        2 * (precision * recall) / (precision + recall)
        if precision is not None and recall is not None and (precision + recall) > 0
        else None
    )
    accuracy = (tp + tn) / (tp + tn + fp + fn) if (tp + tn + fp + fn) > 0 else None

    f_precision = float(f"{precision:.3f}") if precision is not None else "N/A"
    f_recall = float(f"{recall:.3f}") if recall is not None else "N/A"
    f_f1_score = float(f"{f1_score:.3f}") if f1_score is not None else "N/A"

    total_ds = tp + tn + fp + fn

    res = {
        "TP": tp,
        "TN": tn,
        "FP": fp,
        "FN": fn,
        "total": total_ds,
        "P ": f_precision,
        "R ": f_recall,
        "F ": f_f1_score,
        "A ": float(f"{accuracy:.3f}") if accuracy is not None else "N/A",
        # "A ": accuracy,
    }

    return res


def use_table(tp, tn, fp, fn, precision, recall, accuracy):
    total = tp + tn + fp + fn
    total_positive = tp + fn
    total_negative = tn + fp

    table_tn = Table()
    table_tn.add_column("TP", justify="center", style="green")
    table_tn.add_column("TN", justify="center", style="green underline italic")
    table_tn.add_column("FP", justify="center", style="red")
    table_tn.add_column("FN", justify="center", style="red underline italic")
    table_tn.add_column("sum/+/-", justify="right", style="cyan")
    table_tn.add_row(
        f"[bold green]{str(tp)}[/bold green]",
        f"[bold green]{str(tn)}[/bold green]",
        f"[bold red]{str(fp)}[/bold red]",
        f"[bold red]{str(fn)}[/bold red]",
        f"[bold cyan]{str(total)}[/bold cyan]="
        + f"[bold green]{str(total_positive)}[/bold green]+"
        + f"[bold red]{str(total_negative)}[/bold red]",
    )

    table = Table()

    table.add_column("", justify="right", style="cyan", no_wrap=True)
    table.add_column("Formula", justify="center", style="magenta")
    table.add_column("=>", justify="center", style="magenta")
    table.add_column("Value", justify="center", style="magenta")

    table.add_row(
        "",
        f"TP=[bold green]{str(tp)}[/bold green]",
        f"[bold yellow]{str(tp)}[/bold yellow]",
        "",
    )
    table.add_row(
        "Pre",
        "_" * (len(str(tp)) + len(str(fp)) + 6),
        "___",
        f"P=[bold yellow]{precision}[/bold yellow]",
    )
    table.add_row(
        "",
        f"TP=[bold green]{str(tp)}[/bold green] + FP=[bold red underline italic]{str(fp)}[/bold red underline italic]",
        f"[bold yellow]{str((tp + fp))}[/bold yellow]",
        "",
    )

    table.add_section()

    table.add_row(
        "",
        f"TP=[bold green]{str(tp)}[/bold green]",
        f"[bold yellow]{str(tp)}[/bold yellow]",
        "",
    )
    table.add_row("Rec", "______________", "", f"R=[bold yellow]{recall}[/bold yellow]")
    table.add_row(
        "",
        f"TP=[bold green]{str(tp)}[/bold green] + FN=[bold red underline italic]{str(fn)}[/bold red underline italic]",
        f"[bold yellow]{str((tp + fn))}[/bold yellow]",
        "",
    )

    table.add_section()
    # Accuracy
    table.add_row(
        "",
        f"TP=[bold green]{str(tp)}[/bold green] + TN=[bold green underline italic]{str(tn)}[/bold green underline italic]",
        f"[bold yellow]{str((tp + tn))}[/bold yellow]",
        "",
    )
    table.add_row(
        "Acc",
        "______________________________",
        "",
        f"A=[bold yellow]{accuracy}[/bold yellow]",
    )
    table.add_row(
        "",
        f"TP=[bold green]{str(tp)}[/bold green] + TN=[bold green underline italic]{str(tn)}[/bold green underline italic] + FP=[bold red]{str(fp)}[/bold red] + FN=[bold red underline italic]{str(fn)}[/bold red underline italic]",
        f"[bold yellow]{str((tp + tn + fp + fn))}[/bold yellow]",
        "",
    )

    print(table)
    print(table_tn)


def pretty_scores(scores: dict[str, int | float | str | None], draw=True):
    tp, tn, fp, fn = scores["TP"], scores["TN"], scores["FP"], scores["FN"]
    precision, recall, f1_score = scores["P "], scores["R "], scores["F "]
    accuracy = scores["A "]
    # print( f"| TP: {scores['TP']}, TN: {scores['TN']}, FP: {scores['FP']}, FN: {scores['FN']} | ")
    # print(
    #     f"| Precision: {scores['P ']}, Recall: {scores['R ']}, F1 Score: {scores['F ']} |"
    # )
    print(
        f"Precision: [bold cyan]P[/bold cyan] = [bold green]{tp}[/bold green] / ([bold green]{tp}[/bold green] + [bold red]{fp}[/bold red]) = [bold yellow]{precision}[/bold yellow]"
    )
    print(
        f"Recall: [bold cyan]R[/bold cyan] = [bold green]{tp}[/bold green] / ([bold green]{tp}[/bold green] + [bold red]{fn}[/bold red]) = [bold yellow]{recall}[/bold yellow]"
    )
    print(
        f"F1 Score: [bold cyan]F1[/bold cyan] = 2 * ([bold yellow]{precision}[/bold yellow] * [bold yellow]{recall}[/bold yellow]) / ([bold yellow]{precision}[/bold yellow] + [bold yellow]{recall}[/bold yellow]) = [bold magenta]{f1_score}[/bold magenta]"
    )
    print(
        f"Accuracy: [bold cyan]A[/bold cyan] = ([bold green]{tp}[/bold green] + [bold green]{tn}[/bold green]) / ([bold green]{tp}[/bold green] + [bold green]{tn}[/bold green] + [bold red]{fp}[/bold red] + [bold red]{fn}[/bold red]) = [bold yellow]{scores['A ']}[/bold yellow]"
    )

    # use_layout(tp, tn, fp, fn, precision, recall)
    use_table(tp, tn, fp, fn, precision, recall, accuracy)


def main():
    args = parse_args()
    # Example usage
    # tp = 100
    # tn = 50
    # fp = 10
    # fn = 5
    tp = args.tp
    tn = args.tn
    fp = args.fp
    fn = args.fn

    scores = get_scores(tp, tn, fp, fn)
    pretty_scores(scores)


def parse_args():
    parser = argparse.ArgumentParser(
        description="Calculate precision, recall, and F1 score."
    )
    # use positional arguments for tp, tn, fp, fn
    parser.add_argument("tp", type=int, help="True Positives", default=0)
    parser.add_argument("tn", type=int, help="True Negatives", default=0)
    parser.add_argument("fp", type=int, help="False Positives", default=0)
    parser.add_argument("fn", type=int, help="False Negatives", default=0)
    # parser.add_argument("--tp", type=int, required=True, help="True Positives")
    # parser.add_argument("--tn", type=int, required=True, help="True Negatives")
    # parser.add_argument("--fp", type=int, required=True, help="False Positives")
    # parser.add_argument("--fn", type=int, required=True, help="False Negatives")
    return parser.parse_args()


if __name__ == "__main__":
    main()

Day 2605 (17 Feb 2026)

Killing Kubernetes pods without breaking stuff Rancher

To force-kill pods w/o breaking k8s: kubectl delete pod <> --now

Ty KM (https://github.com/kkmarv) for the following, quoting verbatim:

The expected behaviour of kubectl delete heavily depends on its arguments. There are 4 possible but wrongly documented options. I’ve tried to compile different sources into this list:

kubectl delete pod <> will first send a SIGTERM, then a SIGKILL after X seconds (depends on the gracePeriods set by your pod)
kubectl delete pod <> --grace-period=X (X>0) will first send a SIGTERM, then a SIGKILL after X secs
kubectl delete pod <> --now is equivalent to kubectl delete pod <> --grace-period=1

There’s also --force
Where things are possible to go wrong. It immediately deletes the pod from etcd, which will NOT kill the pod. It will too wait for the grace period before sending a SIGKILL. But even then it is not guaranteed that resources will be freed. There is even a warning in kubectl delete --help about this. Only use this when there is no other way to kill a pod.

The force flag also allows to specify --grace-period=0 which will simply default to the grace period defined by the pod, effectively doing nothing since this is also the default behaviour.

Day 2602 (14 Feb 2026)

Passing complex CLI params to Pydantic settings

Settings Management - Pydantic Validation is nice but doesn’t cover everything.

Passing `None` through CLI

Settings Management - Pydantic Validation
None by default b/c cli_avoid_json is the default,[]

it will default to “null” if cli_avoid_json is False, and “None” if cli_avoid_json is True .

Fun fact, works inside complex nested JSON flags as well.

Passing complex JSON-like fields incl. tuples

# Assume this
class MySettingsClass(BaseSettings):
    I_am_a_list_of_tuples: list[tuple[str, str, str | None]] = Field(
        default=FILE_PAIRS, validation_alias=AliasChoices("fp", "file_pairs")
    )
	
FILE_PAIRS = [
    (
        "infer_doc_simplecriteria/pred_doc_simplecriteria.json",
        "merge_gold_simplecriteria/gold_simplecriteria.json",
    ),
    (
        "file2-1",
        "file2-2",
    ),
]

# Passing as a list of lists parses as a tuple
# if done wrong the tuple is parsed as a list of strings
uv run thing --I-am-a-list-of-tuples \
'"merge_gold_simplecriteria/gold_simplecriteria.json",  "infer_doc_simplecriteria/pred_doc_simplecriteria.json"'

Day 2600 (12 Feb 2026)

Pydantic dumping model in json mode with serializable objects

TL;DR pydanticmodel.model_dump(mode=>"json")

Serialization - Pydantic Validation

To dump a Pydantic model to a jsonable dict (that later might be part of another dict that will be dumped to disc through json.dump[s](..)):

settings.model_dump()
>>>
{'ds_pred_path': PosixPath('/home/sh/whatever')}
# json.dump...
json.dumps(settings.model_dump())
>>> TypeError: Object of type PosixPath is not JSON serializable

json.dumps(settings.model_dump(mode="json"))
# works!

Day 2593 (05 Feb 2026)

JQ notes

Shorthand object contsuction

I kept googling for the syntax but could never find it again.

// long: 
'{user: .user, title: .titles}'

// shorthand when key name matches original key name (user==.user)
'{user, title}'
// mix 
'{user, title_name: .title}'

From the man page:

You can use this to select particular fields of an object: if the input is an object with “user”, “title”, “id”, and “content” fields and you just want “user” and “title”, you can write
{user: .user, title: .title}
Because that is so common, there´s a shortcut syntax for it: {user, title}.

The long version allows to operate on they keys:

{id: .idx, metadata: .metadata.original_idx}

Day 2592 (04 Feb 2026)

Simple test to see if VLLM supports structured output

curl 127.0.0.1:8000/v1/chat/completions -s \
  -H "Content-Type: application/json" \
  -d '{
    "model": "/data_pvc2/cache/models--CohereLabs--c4ai-command-r-08-2024/snapshots/96b61ca90ba9a25548d3d4bf68e1938b13506852/",
    "messages": [
      {
        "role": "user",
        "content": "Generate a short analysis in the given JSON schema."
      }
    ],
    "response_format": {
      "type": "json_schema",
      "json_schema": {
        "name": "analysis_schema",
        "schema": {
          "type": "object",
          "properties": {
            "summary": { "type": "string" },
            "importance_score": { "type": "number" },
            "key_points": {
              "type": "array",
              "items": { "type": "string" }
            }
          },
          "required": ["summary", "importance_score", "key_points"],
          "additionalProperties": false
        },
        "strict": true
      }
    }
  }' | jq . > /tmp/out

Sample otuput:

{
  "id": "chatcmpl-9c3932339c7d4e1bace32c24ef6dbada",
  "object": "chat.completion",
  "created": 1770201387,
  "model": "/data_pvc2/cache/models--CohereLabs--c4ai-command-r-08-2024/snapshots/96b61ca90ba9a25548d3d4bf68e1938b13506852/",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "reasoning_content": null,
        "content": "{\n    \"summary\": \"This analysis focuses on the impact of marketing strategies on customer engagement and sales performance for a hypothetical e-commerce company.\",\n    \"importance_score\": 0.85,\n    \"key_points\": [\n        \"Implemented an email marketing campaign with personalized recommendations based on past purchases: Increased open rates by 15% and contributed to a 12% rise in average order value.\",\n        \"Launched an influencer marketing program on social media: Achieved a 20% growth in brand awareness and drove a significant rise in website traffic from social platforms.\",\n        \"Optimised product listings through enhanced SEO techniques: Boosted organic visibility, resulting in a 25% increase in search-driven purchases.\",\n        \"Implemented an automated email nurturing funnel for abandoned carts: Improved cart recovery rate by 10% and increased overall conversions.\",\n        \"Analyzed customer reviews for product feedback and sentiment: Used insights to refine product strategies and enhance customer satisfaction.\"\n    ]\n}",
        "tool_calls": []
      },
      "logprobs": null,
      "finish_reason": "stop",
      "stop_reason": null
    }
  ],
  "usage": {
    "prompt_tokens": 16,
    "total_tokens": 219,
    "completion_tokens": 203,
    "prompt_tokens_details": null
  },
  "prompt_logprobs": null
}

To pretty-print the content it returned:

jq -r .choices[0].message.content /tmp/out | jq .

Simpler OpenAI-compatible JSON mode test

curl 127.0.0.1:8000/v1/chat/completions -s \
  -H "Content-Type: application/json" \
  -d '{
    "model": "/data_pvc2/cache/models--CohereLabs--c4ai-command-r-08-2024/snapshots/96b61ca90ba9a25548d3d4bf68e1938b13506852/",
    "messages": [
      {
        "role": "user",
        "content": "Return a JSON object with keys summary, importance_score (number), key_points (string array)."
      }
    ],
    "response_format": { "type": "json_object" }
  }'

Bonus: literature

Structured output Considered Harmful(tm)
- [2408.02442] Let Me Speak Freely? A Study on the Impact of Format Restrictions on Performance of Large Language Models
- also e.g. Structured outputs create false confidence | Hacker News
Some people STRONGLY disagree
- Say What You Mean: A Response to ‘Let Me Speak Freely’

Day 2551 (25 Dec 2025)

shredding entire directory recursively

security - How do I recursively shred an entire directory tree? - Unix & Linux Stack Exchange:

find <directory_name> -depth -type f -exec shred -v -n 1 -z -u {} \;

Day 2536 (10 Dec 2025)

Indico conference-event-etc. management system

TIL: Indico - Home, allegedly used by ScaDs and CERN, open sourcer

The effortless open-source tool for event organisation, archival and collaboration

Looks really really nice and professional