Jump to content

Template:Infobox model

From Digital Millennium Wiki
Revision as of 00:11, 16 November 2026 by Jdj (talk | contribs) (Pass banner and logo through)

An infobox for AI models: language models, image models, anything with weights. It is Template:Infobox with model fields built in, and every field is optional: empty rows and empty sections are hidden. The example above is made up.

Usage

{{Infobox model
| title           =      <!-- defaults to the page name -->
| banner          =      <!-- wide image across the top, above the title -->
| banner_size     =      <!-- defaults to 400px (shrinks to fit) -->
| banner_caption  =
| logo            =      <!-- logo under the title -->
| logo_size       =      <!-- defaults to 120px -->
| logo_caption    =
| image           =      <!-- e.g. a sample output or diagram; file name without "File:" -->
| image_size      =
| caption         =
<!-- Overview -->
| developer       =      <!-- e.g. [[OpenCerebral]] -->
| family          =      <!-- e.g. Boris -->
| model_type      =      <!-- Base, Instruct, Chat, Reasoning ... -->
| base_model      =      <!-- shown as "Fine-tuned from" -->
| license         =
| status          =      <!-- Training, Released, Preview, Retired ... -->
<!-- Architecture -->
| architecture    =      <!-- e.g. Decoder-only transformer, MHA, RoPE -->
| parameters      =      <!-- 125M, 1.5B ... -->
| layers          =
| context         =      <!-- context length -->
| tokenizer       =      <!-- e.g. 32k BPE -->
<!-- Training -->
| training_tokens =      <!-- 31B, 1.2 trillion ... -->
| tokens_per_param =     <!-- worked out from parameters and training_tokens if left empty -->
| dataset         =
| hardware        =
| training_time   =
| compute         =      <!-- GPU-hours or FLOPs -->
| precision       =      <!-- fp16 mixed, bf16 ... -->
<!-- Benchmarks: score, optional _shots, _ref (a comparison score) and _mean = no -->
| hellaswag       =      | hellaswag_shots     =      | hellaswag_ref     =
| arc_easy        =      | arc_challenge       =
| piqa            =      | winogrande          =      | openbookqa        =
| boolq           =      | sciq                =      | lambada           =
| mmlu            =      | gsm8k               =      | humaneval         =
| bench1          =      | bench1_score        =      <!-- any other benchmark, up to bench30 -->
| bench1_shots    =      | bench1_ref          =
| bench_ref_label =      <!-- column heading for the _ref scores, e.g. GPT-2 XL -->
| bench_mean      =      <!-- "no" hides the mean row -->
| bench_note      =      <!-- metric and harness, e.g. acc_norm, lm-evaluation-harness -->
| bench_source    =      <!-- link to where the numbers come from -->
<!-- Links -->
| weights         =      <!-- e.g. [https://huggingface.co/... Hugging Face] -->
| formats         =      <!-- safetensors, GGUF ... -->
| demo            =
| paper           =
| code            =
| website         =
<!-- History timeline: same events as Template:Infobox -->
| released        =      <!-- release date; 2026, 2026-10 or 2026-10-21 -->
| renamed         =      | renamed_to   =
| event1_date     =      | event1       =      <!-- e.g. weights updated -->
| dissolved       =      <!-- date it was retired -->
| show_age        =      <!-- "no" hides the "... ago" pills -->
<!-- Chronology -->
| series          =      <!-- e.g. Boris models -->
| preceded_by     =
| succeeded_by    =
<!-- extra rows: header26-30, label26-30, data26-30; footer: below -->
}}
  • banner runs across the top of the box, above the title; logo sits under the title. Both are optional and work alongside image.
  • Benchmark scores are percentages; write them as 61.2 or 61.2%. Each gets a small bar.
  • A score that isn't a percentage, such as 18.7 ppl, is shown as written, without a bar, and left out of the mean.
  • With two or more percentage scores, a Mean row is added. Leave a task out of it with hellaswag_mean = no (or bench1_mean = no), or hide the row with bench_mean = no.
  • Give _ref scores (for example a well-known model of the same size) and a Ref. column appears. Name it with bench_ref_label. The reference mean is shown only when every task in the mean has a reference score.
  • _shots = 0 shows zero-shot; 5 shows 5-shot.
  • Tokens / param is worked out from parameters and training_tokens when both are numbers like 125M and 50B.
  • released adds a dated row with how long ago it was. Renames, retirements and other events turn it into a timeline, as in Template:Infobox.
  • Articles using this box are added to Category:Models.