mojzis

• For marimo notebooks: See MARIMO_RULES.md for specific marimo conventions

🚨 CRITICAL MARIMO RULE - NEVER VIOLATE

BEFORE writing ANY marimo cell, check:

1. Are you putting mo.md(), mo.ui.table(), or ANY display function inside if, try, for, while, or with blocks?
2. If YES → STOP! You must prepare content inside control blocks, then display OUTSIDE them.

Template pattern:

willccbb

# train_grpo.py
#
# See https://github.com/willccbb/verifiers for ongoing developments
#
"""
citation:

@misc{brown2025grpodemo,
  title={Granular Format Rewards for Eliciting Mathematical Reasoning Capabilities in Small Language Models},
  author={Brown, William},

bewestphal

Ray Serve Unit Test Examples

1. test_deployment_class.py

Demonstrates how to test a Ray serve deployment class directly without spinning up a Ray cluster.

The key method: func_or_class, allows accessing the underlying class directly where normally, the class would be revised to a Ray serve deployment type due to the @serve.deployment decorator.

2. test_deployment_requests.py

tuulos

import random
from metaflow import FlowSpec, step, S3, Flow, Parameter, profile, kubernetes, conda, conda_base

# change columns according to your schema (or remove column list to load all)
COLUMNS = ['VendorID', 'tpep_pickup_datetime', 'tpep_dropoff_datetime']

# group parquet files as 1GB batches
def shard_data(src, batch_size=1_000_000_000):
    with S3() as s3:
        objs = s3.list_recursive([src])

tuulos

import os
from metaflow import S3

def put_dir(local_root, s3root):
    root = os.path.abspath(local_root)
    objs = []
    for p, _, files in os.walk(root):
        for f in files:
            path = os.path.join(p, f)
            key = os.path.relpath(path, start=root)

vgmoose

Creating a Windows 11 Install USB in 2022

I have been installing Windows for a long time. Does it get easier? I want to say it gets easier, but it seems like there's always some new wrinkle! These instructions are as much a note to my future self as they may be useful to anyone else.

For me, I was not able to get any exfat-based installs, or even any of the GUI helpers to make this process any more straightforward. Maybe on your target Windows / host OS those helpers will work, but the below process (as of current year) is consistent, and not overly complicated.

Overview:

1. Downloading an official ISO image from MS:
   • https://www.microsoft.com/software-download/windows11
2. Formatting the drive (at least 8GB) as GPT, and one FAT-format partition (aka MS-DOS)

• Can use Disks or Disk Utilty for this

miohtama

class LiquidityChanged(TransactionEvent):
    """A sampled liquidity at any moment."""
    __tablename__ = "liquidity"
    delta0 = sa.Column(Int257, nullable=False, index=False)

kklemon

import torch
import torch.distributed as dist
import torch.multiprocessing as mp

from torch.utils.data import IterableDataset, DataLoader


class DistributedIterableDataset(IterableDataset):
    """
    Example implementation of an IterableDataset that handles both multiprocessing (num_workers > 0) 

den-crane

CREATE TABLE test12345 (A Int64) ENGINE = MergeTree  ORDER BY A;

CREATE MATERIALIZED VIEW test12345mv Engine=MergeTree order by A
AS SELECT A FROM test12345;

insert into test12345 select number from numbers(10);

#### maintanance  / load stop

rename table test12345 to test12345_old;

jefftriplett

name: CI

on: [push]

jobs:

  test:
    runs-on: ubuntu-latest

    services: