Threads I started
Threads I started is your own list of discussions, projects, and questions you opened. During the trial, invited members sign in with a link. The template further down helps you write a thread someone can answer.
Threads I started
The list is empty for now.
Beginner
A first experiment, a picture, or a word you want explained
Your list holds discussions you opened: a first experiment, a picture you made, or a question about a word. Keep the draft in a file of your own.
State the problem in the title, as you would to a busy colleague. Include the tool name. "What does temperature mean in this image tool?" tells a helper what you need at a glance. "Help please" does not.
Use four labels. Under What you tried, write the clicks you made or the sentence you typed. What you expected is the result you had in mind. What happened is the actual reply or the exact error, copied as words. The prompt is the full instruction, including any sample text. Describe a picture in words as well as attaching it, so the error can be quoted.
Temperature, when the tool shows it, makes the wording more or less varied. Write that number and the date, because a service can update the model behind the same prompt.
Leave out anything that identifies a living person. The European Commission treats a name, a home address, and a private email as personal data. Use stand-ins such as "Customer A". On its own, a company registration number is not on that list. Keep a named customer's order off the thread.
An empty list means you have published nothing yet. Fill the template and read it as a stranger. When the list is available, each row is a thread you started, with views, saves of your prompt, and a mark on the reply you accept because it worked for you.
Draft my first thread
Help me write a forum post about my own AI experiment. Ask me one question at a time for: the tool and model, whether I used a website, an app, or a program on my computer, what I tried, what I expected, what happened (I will paste the exact reply or error), and the full prompt I sent. Then write a post with a specific title and those four labels. Replace any name, home address, private email, or customer detail with Customer A or another obvious stand-in, and list what you replaced. Keep the temperature value and the date if I mention them.
Intermediate
A workflow, a job prompt, or a retrieval project someone else can run
An intermediate thread shares work another person can run: an automation of two or three tools, a prompt for one job, or a retrieval project. Retrieval-augmented generation (RAG) means the model writes from passages retrieved from documents you supply, and can point at those passages.
Shape the post as a minimal, reproducible example. Minimal: the shortest steps that still show the behavior. Complete: trigger, prompt, model name, and result all sit in the thread. Reproducible: you ran the pasted steps again and saw the same kind of result.
Version the prompt. Mark the first public copy v1 and date it. When a line changes, add v2 and name that line. For a job prompt, name the role, the inputs, the output shape, and the facts the model must copy unchanged. Paste a synthetic input next to the prompt.
For a RAG project, list document types, chunk size if you split files, and the embedding model if you use one. An embedding model turns text into numbers so similar passages can be found. Give one question, the passage you hoped to retrieve, and the passage that came back. Say if the files stayed on your machine or went to a named hosted service.
A private email, a customer name, and a staff record are personal data. Label the example as synthetic. Tick several rows to export one Markdown file, using CommonMark headings, lists, and fenced code, or to follow and unfollow them together.
Make this workflow reproducible
Turn the notes below into a forum thread another person can run. Use a title with the job or workflow, the model, and the date. Include the shortest steps that still show the behavior, the full prompt marked v1, a synthetic input with no real names or private emails, what I expected, and what happened. If this retrieves passages from my documents, add document types, chunk size, embedding model, one question, the passage I hoped to retrieve, and the passage I got. End with a checklist of anything a stranger would still have to guess. Notes: [paste notes]
Expert
A dataset, a fine-tune, a performance note, or a hardware test
An expert thread records a run others can compare: a published dataset, fine-tuned weights, a performance note, or a hardware test. Link the files, the license, and the card at the top.
A model card is the README beside a model: intended use, limits, training settings, datasets, and evaluation. A dataset card is the README beside a dataset: license, language, size, and biases a user should know before training. Name the public base model if you publish an adapter.
LoRA, low-rank adaptation, freezes the original weights and trains small extra matrices inside the network. Hu and colleagues report that, against full fine-tuning of GPT-3 175B with Adam, LoRA reduced trainable parameters by 10,000 times and the GPU memory requirement by 3 times. Report your own run: rank, adapted layers, example count, batch size, sequence length, learning rate, and the metric before and after. A batch is the number of examples processed together. A token is a piece of text the model reads, and sequence length is how many fit in one example.
The GPU holds weights, gradients, and optimizer state. PyTorch keeps freed memory in its own cache, so nvidia-smi can still show it as in use. empty_cache returns unused cached blocks to other programs. Live tensors keep their memory, so a model too large for the card still fails. Paste the out-of-memory text, GPU name, memory figures, command, and library versions. Record power, temperature, and seconds per step with the tool you used.
Mark the thread "needs the machine" when the model does not fit, and keep the failed command in the post.
Write the run report
Draft a forum thread that documents my training or hardware run so someone else can compare it. Ask for any missing item before you write: base model, license, links to the model card and dataset card, LoRA rank and target layers if I used LoRA, example count, batch size, sequence length, learning rate, metric before and after, GPU name, the full out-of-memory text if the run failed, command line, and library versions. Use my numbers only. Where a number is missing, write "not measured". Close with a line I can set to "needs the machine" if the model did not fit in my GPU memory. Notes: [paste notes]
Question template, English and Swedish
Copy this skeleton, fill the brackets, and delete the hints. The same four parts fit a picture, a workflow, or a failed training run.
English
Title: [tool] — [what went wrong or what you want to show], [model name], [date]
What I tried: [steps, in order]
What I expected: [the result you wanted]
What happened: [the actual output or the full error text]
Prompt:
[paste the full prompt]
Setup: [website, app, or local program]; [model]; [settings, including temperature if you set one]; [GPU name and memory, if the work ran on a graphics card]
Example data: [synthetic, or "no personal data"]
Svenska
Titel: [verktyg] — [vad som gick fel eller vad du vill visa], [modellnamn], [datum]
Vad jag försökte: [stegen, i ordning]
Vad jag väntade mig: [resultatet du ville ha]
Vad som hände: [det verkliga svaret eller hela feltexten]
Prompt:
[klistra in hela prompten]
Upplägg: [webb, app eller program på din dator]; [modell]; [inställningar]; [grafikkort och minne, om körningen använde ett grafikkort]
Exempeldata: [påhittad, eller "inga personuppgifter"]
What each row will record
| Column | What it records |
|---|---|
| Views | How many times the thread was opened |
| Saved prompts | How many members saved the prompt from that thread |
| Accepted reply | The reply you marked because it worked in your case |
| Needs the machine | Your flag when the blocker is hardware, such as an out-of-memory stop |
| Follow | Whether new replies should notify you |
Once the list has rows, you can tick several to follow, unfollow, or export together. Export writes the selected threads into one Markdown file: title, the four-part body, setup, and links. CommonMark is the syntax that file follows, so a heading or a code fence still works in another editor. The list contains threads you started. A thread you only answered is not shown here.
Until you have an account, the template above is the part you can finish. Counts begin when a thread is published.
When a thread you started stops because the model does not fit in the graphics memory on your own machine, a shared community DGX Station would be the larger computer that thread could queue for, with the failed command left in the post. About the shared station
Sources
- How do I ask a good question?How to put the specific problem in the title, give context before any code or prompt, and include the steps you already tried.
- How to create a Minimal, Reproducible ExampleWhat minimal, complete, and reproducible mean, and why the expected result and the exact error belong in the post as text.
- CommonMark Spec 0.31.2The plain-text rules for headings, lists, and fenced code blocks used when you export a thread.
- Model CardsWhat to put in the README beside a model: use, limits, training settings, datasets, and evaluation.
- Dataset CardsWhat to put in the README beside a dataset: license, language, size, and biases that affect reuse.
- CUDA semantics — Memory managementWhy PyTorch can still look busy in nvidia-smi, what empty_cache releases, and why memory held by live tensors stays allocated.
- LoRA: Low-Rank Adaptation of Large Language ModelsThe paper that defines LoRA and reports the GPT-3 175B comparison on trainable parameters and GPU memory.
- Data protection explainedThe European Commission's examples of personal data, including a name, a home address, and a private email, and the note that a company registration number on its own is outside that category.