It does, but it's very poorly documented and quite unstable (on purpose i think). What the other commenter said about the VSCode BYOK seems to be the more reliable way.
I tried adding a Foundry LLM as Github Copilot custom model and failed miserably. But with VSCode BYOK (and Github Copilot as the interfact) i did get it working, and i can now use Deepseek V4 Flash with Copilot.
I don't think youre missing out necessarily. I would riot if i couldn't use Pycharm anymore, for big python projects just nothing beats it right now.
I do use VSCode too, but mostly for quick scripting or non-programming projects. and even then i installed a bunch of extensions to make it more like Pycharm.
Not sure, i can't run it since i can't install Node.js on my work environment. What is your experience with Context7 like?
As for GitMCP: I think the url fetching tool of the docs it does is not great, but the code searching tool is quite good. Regardless, i remain open to alternatives, not stuck to this yet.
I've been using GitMCP.io + Github Copilot for this problem specifically (AI assistant + accurate docs). The downside is that you need to add a separate MCP server for each repository, but the qualitative difference in agent mode is incomparable.
I used it recently to do a major refactor and upgrade to MLFlow version 3.0. Their documentation is a horrid mess right now, but the MCP server made it a breeze because i could just query the assistant to browse their codebase. Would have taken me hours extra myself.
Good to see positive reception to feedback! Sorry if my message came out as condescending, was not the intent. I recommend reading this piece on metrics https://openforecast.org/wp-content/uploads/2024/07/Svetunko.... It's easy to grasp, yet it contains great tips.
Short answer: i use multiple metrics, never rely on just 1 metric.
Long answer: Is the metric for people with subject-matter knowledge? Then (Weighted)RMSSE, or the MASE alternative for a median forecast. WRMSSE is is very nice, it can deal with zeroes, is scale-invariant and symmetrical in penalizing under/over-forecasting.
The above metrics are completely uninterpretable to people outside of the forecasting sphere though. For those cases i tend to just stick with raw errors; if a percentage metric is really necessary then a Weighted MAPE/RMSE, the weighing is still graspable for most, and it doesn't explode with zeroes.
I've also been exploring FVA (Forecast Value Added), compared against a second decent forecast. FVA is very intuitive, if your base-measures are reliable at least. Aside from that i always look at forecast plots. It's tedious but they often tell you a lot that gets lost in the numbers.
RMSLE i havent used much. From what i read it looks interesting, though more for very specific scenarios (many outliers, high variance, nonlinear data?)
Look i'm optimistic about time-series foundation models too, but this post is hard to take seriously when the test is so flawed:
- Forward filling missing short periods of missing values. Why keep this in when you explictly mention this is not normal? Either remove it all or don't impute anything
- Claiming superiority over classic models and then not mentioning any in the results table
- Or let's not forget, the cardinal sin of using MAPE as an evaluation metric
Good list. Some of these i knew already, but the typing overloading and keyword/positional-only arguments were new to me.
One personal favorite of mine is __all__ for use in __init__.py files. It specifies which items are imported whenever uses from x import *. Especially useful when you have other people working on your codebase with the tendency to always import everything, which is rarely a good idea.
What a great article, i always like how much Anthropic focuses on explainability, something vastly ignored by most. The multi-step reasoning section is especially good food for thought.
With good reason honestly. They take all the best practices from existing tooling we had, discard the bad, and make it run blazingly fast.
Ruff for me meant i could turn 4 pre-commit hooks (which you have to configure to be compatible with each other too) into just 1, and i no longer dread the "run Pylint and take a coffee break" moment.
I jumped ship to UV recently. Though i was skeptical at first i don't regret it. It makes dependency management less of a chore, and just something i can quickly do now. Switching from Poetry was easy for me too, only package i had issues with was pytorch, but that just required some different toml syntax.
The comparison in general is quite poor. Where is a basic model list? You know, one of the biggest choices when forecasting...
Also, when saying "Nixtla" do they mean TimeGPT or one of their other libraries? Because the former definitely supports exogenous regressors, while they say it doesn't.
I personally prefer Darts because it's very user-friendly, and the devs are responsive,
I tried adding a Foundry LLM as Github Copilot custom model and failed miserably. But with VSCode BYOK (and Github Copilot as the interfact) i did get it working, and i can now use Deepseek V4 Flash with Copilot.