How the disenchantment page works
The /market/disenchantment page lists every foundation model with at least three abliterated descendants - currently 324 bases, ranked by DAI. This article walks through the columns, the ACTIVE / INACTIVE badge, the per-row 24-month sparkline, the click-to-expand full history chart, and the "N forks" link into the catalog.
The page is a sortable, searchable table with one row per foundation model. Columns: base ID, base downloads, descendants count, total descendant downloads, DAI, and a 24-month sparkline. Each row has an ACTIVE / INACTIVE badge based on whether descendant downloads have grown or contracted over the last three months. Clicking a row expands it to show the full 24-month history as two lines: base downloads and descendants downloads. Clicking "N forks" opens the catalog filtered to that base's entire tree.
- What each column on the table means and how it is computed
- The ACTIVE / INACTIVE badge - the exact 3-month trend rule
- The per-row sparkline - 24 months of DAI compressed into 100 pixels
- The click-to-expand chart - base line vs descendants line, why two colors
- The 'N forks' link - how it connects the page to the catalog
- How the page differs from the DAI wiki article and the market Signals block
What the page is
The /market/disenchantment page is the full detail view for the Disenchantment Index. It lists every foundation model in our tracked set that has at least three abliterated descendants, ranked by DAI. Currently 324 rows.
The market page (/market) shows a top-10 preview of the same data in its Indices tab. This page shows all of them, sortable, searchable, with per-row visualization. It is the "dig in" destination when the market summary is not enough.
Columns, one at a time
Base ID
The HuggingFace model identifier of the foundation model - the root of the fork tree we are analyzing. Clickable link to the model's HuggingFace page.
Base downloads
Lifetime downloads of the base model itself, from the latest monthly snapshot. This is the denominator of DAI: how big the audience for the original is.
Descendants
Count of abliterated models in the base's recursive fork tree at any depth. Direct forks, forks of forks, quantizations of forks, merges - everything descending from this base, filtered to only those in our abliterated catalog. This count matches what you get when you click "N forks" on the same row.
Descendant downloads
Sum of lifetime downloads across every model in that descendant set. This is the numerator of DAI: how big the audience for the abliterated versions is in aggregate.
DAI
The disenchantment reading for this base. Formula: descendant downloads divided by base downloads, expressed as a percent. Values above 100% are not errors - they mean the abliterated audience is larger than the base audience. The full mathematical treatment lives in the DAI wiki article.
Sparkline
A compressed 24-month history of DAI for this base, drawn in the space of one table cell. Higher trend line = DAI rising for this specific base. Flat line = stable. Falling line = descendants growing slower than the base. Hover for a full chart (see next section).
The ACTIVE / INACTIVE badge
Every row carries a badge in one of two states:
- ACTIVE - descendant downloads have grown or held stable over the last three months
- INACTIVE - descendant downloads have dropped by more than 30% over the last three months
Currently the table holds 0 ACTIVE bases and 324 INACTIVE. The badge is derived from the descendants' downloads trajectory across the three latest snapshots, not from the DAI reading itself. A base can have a high DAI while being INACTIVE - meaning its abliterated tree once captured a lot of audience but the audience has since moved on.
Why 30%. The threshold is a judgment call. Small enough to catch real audience departure; large enough to filter month-to-month noise from snapshot to snapshot. If a descendant tree lost more than 30% of its monthly downloads over a quarter, that is not noise - that is the audience leaving. The 30% cutoff is the same order of magnitude as other quarterly-decline flags used in analytics.
Sparklines - per-row DAI history in 24 months
The rightmost column of each row shows a mini line chart - the base's DAI reading at each of the 24 monthly snapshots we hold. Purpose: to show at a glance whether this base's abliterated share has grown, shrunk, or stayed flat over the covered window.
Sparklines are auto-scaled to their own y-axis range - the point is shape, not absolute magnitude. Two sparklines side by side are showing local movement each in their own coordinate space; they are not directly comparable in height.
Click-to-expand - the full history chart
Clicking any row expands it in place and reveals a full-size chart showing both signals as separate lines:
- Yellow line - base downloads. How the audience for the original model has moved month by month.
- Purple line - descendant downloads. How the aggregate audience for its abliterated tree has moved month by month.
Two-line format is the clearest way to see why DAI moved. If the purple line rose while yellow was flat, the abliterated audience grew and pushed DAI up on its own. If both rose together, the whole family got popular. If yellow fell while purple held, the base was abandoned but descendants stayed - DAI rose because the denominator shrank.
Clicking the row again collapses it. Multiple rows can be expanded at once for side-by-side comparison.
The "N forks" link - going into the catalog
Each row has a "N forks" link where N is the descendant count. Clicking it opens the main catalog at /models?base=<base_id>, filtered to show every descendant of this base at any tree depth.
The count on the button matches the count in the catalog filter - both come from the same recursive fork-tree query. If a base row says "20 forks," you will see 20 rows in the filtered catalog, one per model. From there you can drill into individual model cards, sort by downloads, filter by method, or export the list.
How the page differs from related resources
Three resources describe overlapping but distinct things:
- This page (/market/disenchantment) - the operational data: every base, every DAI, every trajectory. Interactive.
- The DAI wiki article (/wiki/frontiers/disenchantment-index) - the methodology and philosophy: what DAI means, how it is calculated, why it is named after Weber. Static reference.
- The Indices tab preview on the market page (/market#indices) - the top-10 summary, plus flagship number and two 24-month charts. Overview.
Reading them in order - market preview to wiki article to detail page - is the intended flow for someone new to the metric. Detail page first is fine too if you already know what DAI is.
How the page updates
Every night at 05:00 UTC an automation pulls the latest HuggingFace snapshot, recomputes descendant trees, updates the DAI history tables, and rebuilds the Astro site. The numbers you see on this page and its parent Indices block are never more than 24 hours old. Full pipeline description in the DAI wiki article.
Where to read more
- Disenchantment Index (DAI) - full methodology and philosophy of the metric itself
- How the Indices tab works - parent article covering all four flagship indices
- Every number on the market page, explained - master guide across all three tabs
- The page itself