ncoghlan

(I wanted an initial prompt to check if LM Studio would run this model on my machine, and finding out more about it seemed like a reasonable place to start)

Prompt

Tell me about your training environment and the resulting limitations on your capabilities.

Response

Okay, let's break down my training and its impact on what I can (and can't) do. I'm Gemma, an open-weights AI assistant created by the Gemma team at Google DeepMind. Here's a detailed look:

ncoghlan

>>> class C:
...    def example(self): pass
... 
>>> import pickle
>>> data = pickle.dumps(C.example)
>>> pickle.loads(C.example)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: a bytes-like object is required, not 'function'
>>> pickle.loads(data)

ncoghlan

Using ?it as a statement local symbolic reference

Core concept

• (?it=expr) is a new atomic expression for an "it reference binding"
• subsequent subexpressions (in execution order) can reference the bound subexpression using ?it (an "it reference")
• ?it is reset between statements, including before entering the suite within a compound statement (if you want a persistent binding, use a named variable)

ncoghlan

Copyright 2017 Nicholas Coghlan. All rights reserved.

Redistribution and use in source and binary forms, with or without modification, are
permitted provided that the following conditions are met:

   1. Redistributions of source code must retain the above copyright notice, this list of
      conditions and the following disclaimer.

   2. Redistributions in binary form must reproduce the above copyright notice, this list
      of conditions and the following disclaimer in the documentation and/or other materials

ncoghlan

# Prompted by this article about implementing "Round half away from zero" in Go
# https://www.cockroachlabs.com/blog/rounding-implementations-in-go/

# Python uses round-half-even by default to reduce statistical bias,
# but round-half-away-from-zero can be implemented more efficiently
# (since it doesn't need to examine all the bits of the number)

# I figured the folks doing computational analysis in Python would
# have made sure this was handled correctly years ago, but it never
# hurts to check that kind of assumption :)

ncoghlan

PEP: 538 Title: Coercing the legacy C locale to a UTF-8 based locale Version: $Revision$ Last-Modified: $Date$ Author: Nick Coghlan <[email protected]> BDFL-Delegate: INADA Naoki Status: Draft Type: Standards Track Content-Type: text/x-rst Created: 28-Dec-2016

ncoghlan

# Using even one additional keyword arg forces CPython to make an additional copy of the entire original keyword dictionary
$ python3 -m timeit -s "kwds = {str(k):v for k, v in enumerate(range(1000))}" -s "def f(**kwargs): return kwargs" "f(**kwds)"
10000 loops, best of 3: 50.5 usec per loop
$ python3 -m timeit -s "kwds = {str(k):v for k, v in enumerate(range(1000))}" -s "def f(**kwargs): return kwargs" "f(other=1, **kwds)"
10000 loops, best of 3: 82.4 usec per loop


# Using a non-dict subclass forces CPython to copy it to a real dictionary as part of the call
$ python3 -m timeit -s "import collections; kwds = collections.UserDict((str(k), v) for k, v in enumerate(range(1000)))" -s "def f(**kwargs): return kwargs" "f(**kwds)"
1000 loops, best of 3: 573 usec per loop

ncoghlan

>>> class Madness(dict):
...     def __setitem__(self, attr, value):
...         if isinstance(value, type):
...             value.__name__ = attr
...         dict.__setitem__(self, attr, value)
... 
>>> code = """\
... class Example(object): pass
... NewName = Example
... print(Example.__name__)

ncoghlan

class MixedTypeKey:
    """Sort key for mixed type lists in Python 3

    >>> sorted([None, ..., 1, 1j, "", (), {}, []])
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
    TypeError: unorderable types: ellipsis() < NoneType()
    >>> sorted([None, ..., 1, 1j, "", (), {}, []], key=MixedTypeKey)
    [None, Ellipsis, 1, 1j, '', (), {}, []]
    """

ncoghlan

Github won't let me comment on https://github.com/pculture/unisubs/commit/e41f74fa5a17573b08001cc1431b615abca930ba

so just saving this here for now.

unisubs is hitting the double import problem that is caused by Django's old default project layout and management script, which creates a broken Python import configuration. Django's default behaviour has been fixed (https://docs.djangoproject.com/en/dev/releases/1.4/#updated-default-project-layout-and-manage-py) but projects using the old layout may still see the issue.

The only reliable solution is to fix the layout to avoid the broken configuration that allows the same module to be imported under two different names. The simplest way to do this is to:
1. Remove any sys.path manipulation from manage.py
2. Ensure that manage.py is *not* inside a Python package (there should *not* be an adjacent __init__.py file)
3. Adjust imports as needed for any changed made in steps 1 and 2