Skip to content

Query Examples

Rohan Vinaik edited this page Mar 7, 2026 · 2 revisions

Getting Value Fast · Chapter 2 of 2 · Prerequisites: Getting Started · ~4 min read

A cookbook of questions ModelAtlas can answer that HuggingFace cannot — and the structural reasons why.


Problem Space

Search engines match keywords. Filter UIs match columns. Neither can express: "models like this one but smaller and more code-focused." That query requires understanding proximity in multiple dimensions simultaneously — a navigation problem, not a search problem.

Every example below shows a question, how the LLM decomposes it into navigate_models parameters, and what structural features produce the answer.


Navigation Queries

"Small code model with tool-calling for a MacBook"

efficiency:    -1  (smaller than mainstream)
capability:    +1  (specialized)
compatibility: +1  (platform-specific)
require:  [code-generation]
prefer:   [tool-calling, Apple-Silicon-native, instruction-following]

Why HF can't answer this: No field combines size direction, capability specialization, and platform compatibility. You'd need to manually intersect task filters, guess at parameter counts, and grep model cards for "Apple Silicon."

What ModelAtlas does: Bank alignment scores penalize models on the wrong side of any dimension. The require filter uses SQL pre-filtering. IDF weighting ensures rare anchors like Apple-Silicon-native count more than ubiquitous ones like decoder-only.

"Models architecturally similar to Mamba but instruction-tuned"

architecture:  +1  (novel architectures)
require:  [instruction-following]
prefer:   [mamba, ssm]

Why HF can't answer this: HuggingFace has no "architecture" axis. You can filter by model_type, but that's a flat string — mamba and rwkv don't know they're related (both are SSMs, both are non-transformer).

What ModelAtlas does: The ARCHITECTURE bank places Mamba at positive signed position (novel relative to the transformer zero state). SSM models cluster nearby. The require filter ensures only instruction-tuned variants return.

"What's the base model for this GGUF quantization?"

similar_to: "TheBloke/Llama-3.1-8B-Instruct-GGUF"
lineage:    -1  (toward base/foundational)

Why HF can't answer this: GGUF files are uploaded by quantizers, not the original authors. The provenance chain (quantized_from → fine_tuned_from → base model) is implicit in naming conventions and model card text, not in any queryable field.

What ModelAtlas does: Explicit quantized_from and fine_tuned_from links in the network. The LINEAGE bank encodes derivation depth. Seeding from the GGUF model and moving negative in LINEAGE traverses back toward the base.

"Trending medical models with strong benchmarks"

domain:   +1  (specialized)
quality:  +1  (above mainstream)
require:  [medical-domain]
prefer:   [trending, high-mmlu, instruction-following]

Why HF can't answer this: "Trending" is a sort order on HF, not a filterable property. "Medical" is a tag some authors apply. "Strong benchmarks" requires parsing model cards or external leaderboards.

What ModelAtlas does: The QUALITY bank tracks trend signals. The DOMAIN bank encodes specialization depth. Benchmark-derived anchors like high-mmlu are extracted during classification and linked with provenance.


Comparison Queries

"How do Llama 3.1 8B and Qwen 2.5 7B differ?"

Uses compare_models — set operations on anchor sets plus per-bank position deltas.

Returns:

  • Shared anchors: What they have in common (decoder-only, instruction-following, 7B-class)
  • Unique to each: What distinguishes them (family, training approach, specific capabilities)
  • Bank deltas: Where they sit differently across dimensions
  • Jaccard similarity: Quantified structural overlap

"What models are most similar to CodeLlama-34B?"

Uses similar_to — IDF-weighted Jaccard between the seed model's anchor set and every other model's. Returns the structurally nearest neighbors with the specific anchors that drove the similarity score.


Discovery Queries

"What kinds of models exist in the speech domain?"

domain: +1
prefer: [speech-domain, audio-domain]

Returns models clustered by speech-related anchors. The anchor sets on the results reveal the sub-structure: text-to-speech, speech-recognition, voice-cloning, audio-generation — categories that emerge from the data rather than being imposed by a taxonomy.

"Show me the frontier of efficient reasoning models"

efficiency: -1  (small)
capability: +1  (specialized)
training:   +1  (advanced training)
prefer: [reasoning, chain-of-thought, distilled]

The combination of opposing directions (small BUT capable, advanced training) naturally surfaces models that punch above their weight class — the interesting part of the design space.


What These Examples Have in Common

Every query above requires simultaneous reasoning across multiple dimensions — something keyword search cannot do and filter UIs cannot express. The signed hierarchy model converts each dimension into arithmetic, and the multiplicative scoring ensures no dimension is ignored.


Related Concepts


← Getting Started · System Overview →

Clone this wiki locally