How the Signals tab works

Why a separate Signals tab exists

The Overview tab shows the state of the market at a glance - counts, families, momentum, evolution chains. The Indices tab shows flagship metrics that track evolution over time. The Signals tab is different: each block reveals a structural property of the market that is not immediately visible from raw counts.

Concentration is not visible from counting models. Activity is not visible from listing them. Originality of methodology is not visible from tagging families. Every Signals block answers a question that requires stepping back from the numbers and looking at the shape they form.

1. Pareto concentration - the 1 percent rule

In a healthy competitive market, the top 1 percent of products captures 20-30 percent of demand. In this niche, the number is far higher.

Current reading: the top 1% of models (N/A models) capture N/A% of all downloads in the catalog. The top 5% capture around 95%. The remaining 95% of the catalog fights over the last 5% of attention.

This is an extreme long-tail distribution - much more skewed than typical software markets, closer to what social platforms show for viral content or what publishing shows for bestsellers. It has direct consequences for two audiences: authors looking to build a following (the top spots are essentially locked), and consumers deciding what to download (the top 1% is where the vast majority of quality-checked work sits).

Formula: sort every model by lifetime downloads descending. Compute the cumulative sum. The share captured by the top N% is the cumulative sum at rank N% divided by the total. Standard Pareto analysis.

2. Activity breakdown - how much of the catalog is still alive

Cataloged does not mean used. Every model in this niche has a live counter from HuggingFace showing downloads over the last 30 days. The activity block classifies the catalog into six buckets by that counter, then aggregates them into three hero cells for readability.

• Actually active: 2,773 models (15.7%) get 1,000+ downloads per month. These are the workhorses of the ecosystem.
• Have real traffic: 9,184 models (51.9%) get 100+ downloads per month. Any use above this threshold is not incidental.
• Effectively forgotten: 8,495 models (48.1%) get less than 100 per month. The bulk of the catalog sits here.

The full six-bucket distribution (10K+, 1K-10K, 100-1K, 10-100, 1-10, 0) appears below the hero cells with horizontal bars. Roughly 8% of the catalog gets zero downloads in a given 30-day window - not dead in the sense of removed, but functionally invisible.

This block does what raw catalog counts cannot: it separates existence from use. A number like "16,000 indexed models" sounds like a lot; the activity breakdown reveals that only a fraction of that number is doing real work at any given moment.

3. Originality - how much is fresh work

Not every model in the catalog represents new abliteration research. Many are repackagings of existing work - a GGUF quantization of someone else's abliteration, or a merge of already-abliterated models. The originality block splits the catalog into four categories to make this distinction explicit.

• Original abliteration or fine-tune (9,875, roughly 54%): actual weight surgery or preference training. M1, M3, M4, M-uncensored. This is where new abliterated models come from.
• Repackaged (GGUF conversions) (roughly 32%): M8. Existing abliterated models re-quantized for llama.cpp / Ollama / LM Studio. Necessary work, but no new abliteration.
• Merged (mergekit) (roughly 5%): M5. Weight arithmetic combining several already-abliterated models. Second-order production.
• Not yet classified (roughly 9%): metadata insufficient for the classifier. Coverage improves as we index more.

Below the four hero cells, a table shows the full breakdown by method (M1 through M8, plus M-uncensored and unclassified). For each method it shows share of catalog and share of downloads, which lets you spot mismatches - methods that produce many models but attract few downloads, or vice versa.

4. Demand-side inequality (Gini)

The Overview tab shows a supply-side Gini - how unequal is production, or how much do a few authors publish. This block shows the demand-side Gini - how unequal are downloads, or how much do a few models absorb attention. They usually differ: producers concentrate around a few prolific pipelines, but demand can concentrate even more sharply on a handful of proven models.

Current demand Gini: 0.943 - extreme concentration. Computed across 16,244 models representing 52,375,495 lifetime downloads. For reference: US household income Gini is roughly 0.48; values above 0.9 are rare outside heavy-tailed digital markets like open-source packages or viral content.

Formula: sort every model's lifetime downloads ascending. Let x_i be the downloads at rank i, and n the total number of models. Then Gini = (sum over i of (2i - n - 1) * x_i) / (n * sum(x_i)). Standard economics-textbook version, applied to downloads rather than income.

The Gini number and the Pareto number above answer overlapping but not identical questions. Pareto asks how much the top slice captures; Gini asks how unequal the whole distribution is. A market can have a high Pareto reading (top slice captures a lot) but a moderate Gini (the long tail is fairly even). This niche shows both - high concentration at the top and extreme unevenness overall.

5. Context length distribution

Every abliterated model with GGUF or safetensors metadata declares a context length - the maximum token window it can handle. The distribution is unusual.

Context sizes cluster tightly at powers of two: 4K, 32K, 128K, and 256K account for the bulk of the catalog. Then there is a real gap between 256K and 1M - almost no models sit in that range. A single spike appears at 1M (2^20 tokens, the nominal ceiling for most families). Beyond that, a handful of Llama-4-Scout variants reach all the way to 10M.

The gap between 256K and 1M is genuine market intelligence. It suggests that either (a) intermediate context sizes are technically awkward for the transformer architectures in play, or (b) the community has settled on a two-tier standard (256K for most work, 1M for niche uses) and skips intermediate values. Either interpretation is a factual statement about how the market has organized itself around context.

6. Method firsts

For each classified abliteration method, the earliest model on record. Not "who invented the technique" - that is often academic - but who published the first release we can classify with confidence.

Currently tracked: methods M1 through M8 plus M-uncensored, M5-frankenstein. The earliest release in each family sits in the block as a card with author, model, method code, date, and download count. Reading the dates left to right gives a rough sense of the field's chronology: M8 repackaging appeared first (September 2023), M5 mergekit-style work followed (December 2023), M1 direct removal in mid-2024, and so on.

This block anchors historical claims. When a wiki article or a press piece says "abliteration first appeared in 2024," what that really means is "the first M1 abliteration in our catalog appeared in mid-2024" - a specific model by a specific author on a specific date. The method-firsts block makes those anchors clickable.

7. Disappeared models - the deletion timeline

Nowhere else public tracks this. Over the 24-month window we cover, 2,860 models that once appeared in this niche have vanished from HuggingFace - roughly 15% of everything ever indexed. Average lifespan before deletion: about 8 months.

Some had over a million lifetime downloads before disappearing. Reasons vary: author account deletions, HuggingFace policy takedowns, model card violations, or private choices by the author to withdraw. The catalog does not attempt to reconstruct why - only to record that a model existed at time T, was last seen at time T+N, and then disappeared.

The Signals block shows a monthly timeline as a bar chart (disappearances per month), plus a short list of the most significant deletions by lifetime downloads. The full 2,860-row table with sort and search lives at /market/deletion-history.

Formula: the historical_deletions table is rebuilt each night. A model is considered "deleted" if it appeared in at least one earlier HuggingFace snapshot but is absent from the latest snapshot. Its "last seen" month is the last snapshot it appeared in; its "peak downloads" is the maximum download count observed across all snapshots where it appeared. Model catalog metadata (files, size, family) is preserved from the last snapshot in which the model was live.

8. Refusal direction extraction methods

How the refusal direction was found for each model where we can determine it. The M1-M8 classifier tags what was done to the weights; this block shows the geometric method behind that operation.

Seven methods are documented in the academic literature. Only two are visible in production: Difference-of-means (used by the vast majority of released abliterations) and huihui-ai's layer-band extraction (used by that author's models specifically). The other five - Affine Concept Editing (ACE), Contrastive Activation Addition, PCA/SVD on activation differences, Recursive Feature Machines with AGOP, Self-Organizing Maps - exist on paper but have not been observed in a released model.

Theory has run ahead of practice. The distribution is one useful angle on the state of the field: rich theoretical vocabulary, narrow production practice. This mismatch is itself signal - it says that most abliterations in the wild use one of the two simplest available techniques, even though the literature offers many more sophisticated ones.

How the Signals blocks fit together

Each block answers a different structural question. Read together, they paint a coherent picture of the market:

• Attention is extremely concentrated (Pareto)
• Most catalog entries are effectively inactive (Activity)
• A large minority of production is repackaging, not new work (Originality)
• Downloads are more unequal than incomes in most countries (Demand Gini)
• Technical choices cluster around convention (Context length)
• Methods entered production in a traceable chronology (Method firsts)
• A significant share of published work disappears over time (Deletion timeline)
• Practice uses far fewer techniques than the literature offers (Extraction methods)

A market that is extremely concentrated, largely inactive, heavily repackaged, unequally downloaded, technically conservative, chronologically traceable, prone to deletion, and methodologically narrow. Each Signals block contributes one adjective to that description.

Where to read more

• How market metrics work - the master guide covering all three tabs
• Disenchantment Index (DAI) - the flagship metric from the Indices tab
• The fork graph - how family trees are constructed from base-model declarations
• Methods overview (M1-M8) - the classification behind the Originality block
• What is refusal direction - the underlying concept behind the Extraction methods block