juiceb0xc0de/b-sides-index

B-Sides Index

The search index behind B-Sides — deep search for the Hugging Face models that never charted.

---

Hugging Face search: great at finding a model you already know the name of. It's useless for the other questions. What's out there that everyone let slide by unnoticed? Which unassuming repo with 3 downloads is actually a university lab's RL experiment? What got uploaded in April 2025 that trains with GRPO on a qwen3 base and supports flash_attention implementation?

So I built my own search engine.

228,610 text-generation models to date, embedded and filed. Every one of them a B-side: the chart-toppers are gone, the quant reissues are gone, and what's left is the long tail where the actual interesting work hides.

What's in the crate

Two files. Row i in one is row i in the other.

• embeddings.f16.npy — a 228,610 x 1536 fp16 matrix, one row per model, pre-normalized so cosine similarity is a plain dot product. Embeddings are OpenAI text-embedding-3-small over each model's metadata: id, author, upload month, tags, download and like counts.
• meta.sqlite — the filterable side of the house. One row per model: upload month, creation timestamp, tags, license, languages, and the lineage columns. Which base model it descends from and how (48,771 finetunes, 8,303 adapters, 7,718 merges).

What got thrown out

The index is defined as much by what's not in it.

Official org releases - Qwen, Meta, Google, DeepSeek, Microsoft, the whole A-side - filtered out, unless the repo ships custom code or a paper. You don't need my index to find Qwen3-8B.

Quant reissues and format conversions - GGUF, AWQ, GPTQ, EXL2, MLX, the endless -4bit dumps - gone. This is a model directory, not a quant directory. One representative pressing per record.

What survives is finetunes under unassuming names, beta checkpoints, merges somebody actually thought about, one-off architectures, and repos with 0 downloads that deserve better. Popularity is a smokescreen out here. Some of the best records are found in unopened at an estate sale forgotten by everyone including the owner.

The pressing

This is the 2025 pressing: models by creation date, 2022-03 through 2025-12. One honest wrinkle — the 2022-03 bucket is inflated, because Hugging Face stamped that date on every repo that predates creation tracking. Treat that month as "date unknown," not as a March 2022 gold rush.

The enrichment columns in meta.sqlite - parameter counts, full config.json, context length, and which modules each repo's custom code actually imports (flash_attn, causal_conv1d, the works) — are being backfilled by a collection pass over all 228k repos right now. They land in a future revision of this dataset. The schema is already there.

Caveats, quickly

Embeddings cover metadata, not model cards. The embedded text is id, tags, and stats — semantic search finds naming and tag vocabulary, not README prose. The richer embedding pass is in progress.

Downloads and likes are snapshots. Frozen at collection time. Fine for coarse sorting, stale for anything else. This caveat is irrelevant for the most interesting discoveries I've made so far.

The slop filter is regex, not judgment. A name-based filter catches most quant dumps. I'm just a bartender, use what I design with your a sense of caution and curiosity.

One embedding model. Query vectors must come from text-embedding-3-small or the geometry means nothing. The embedding model is provided with the Hugging Face space.

Footnote: run it yourself

-- what was quietly training with GRPO on qwen3 last spring?
sqlite3 meta.sqlite "
  SELECT model_id, month, downloads
  FROM models
  WHERE tags LIKE '%\"qwen3\"%' AND tags LIKE '%\"grpo\"%'
    AND month BETWEEN '2025-04' AND '2025-06'
  ORDER BY downloads ASC LIMIT 20;
"

Sort ascending on purpose. The good stuff is at the bottom of the crate.

Go dig something.