This commit is contained in:
76
packages/agent/src/model-family.ts
Normal file
76
packages/agent/src/model-family.ts
Normal file
@@ -0,0 +1,76 @@
|
||||
/**
|
||||
* Model family classifier — orthogonal to ModelTier.
|
||||
*
|
||||
* Tier (small / mid / frontier) drives scoring profile, recall limit, and
|
||||
* scaffold gating. Family drives scaffold-STYLE selection (compression vs
|
||||
* expansion).
|
||||
*
|
||||
* v4 findings motivating this split:
|
||||
* - F > E on 5/6 scenarios: Opus 4.6 + PA beats Opus 4.6 without PA. Claude
|
||||
* benefits from compression scaffolds at frontier tier.
|
||||
* - Qwen3-30B-A3B (instruct variant) showed +26.7pp on compare scenario with
|
||||
* compression scaffold.
|
||||
* - Gemma 4 31B + 26B MoE regressed on 5/6 scenarios with compression
|
||||
* scaffold.
|
||||
*
|
||||
* Working hypothesis (v5): reasoning-capable families gain from explicit
|
||||
* structure; dense instruction-tuned families without reasoning bias are hurt
|
||||
* by compression-style scaffolds because compression fights their natural
|
||||
* elaboration style. Gemma may still benefit from an EXPANSION scaffold.
|
||||
*
|
||||
* See PromptAssembler v5 brief §7.1.
|
||||
*/
|
||||
|
||||
export type ModelFamily =
|
||||
| 'claude'
|
||||
| 'gemma'
|
||||
| 'qwen-reasoning'
|
||||
| 'qwen-instruction'
|
||||
| 'llama'
|
||||
| 'other';
|
||||
|
||||
/** Reasoning markers that disambiguate qwen-reasoning from qwen-instruction. */
|
||||
const QWEN_REASONING_MARKERS = ['thinking', 'reasoner', 'reasoning'] as const;
|
||||
|
||||
/**
|
||||
* Map a model identifier to its training-lineage family.
|
||||
*
|
||||
* Accepts bare slugs (`claude-opus-4-7`), provider-prefixed slugs
|
||||
* (`google/gemma-4-31b-it`), and nested LiteLLM/OpenRouter slugs
|
||||
* (`openrouter/qwen/qwen3-30b-a3b-thinking-2507`). Case-insensitive.
|
||||
*
|
||||
* Unknown models default to `other` (conservative — callers should treat
|
||||
* `other` as "don't assume reasoning behavior").
|
||||
*/
|
||||
export function familyForModel(model: string): ModelFamily {
|
||||
const normalized = model.trim().toLowerCase();
|
||||
if (!normalized) return 'other';
|
||||
|
||||
// Strip provider prefix(es): keep only the last slug segment.
|
||||
// Handles: "google/gemma-4-31b-it", "qwen/qwen3-...", "openrouter/qwen/qwen3-..."
|
||||
const bare = normalized.includes('/')
|
||||
? normalized.slice(normalized.lastIndexOf('/') + 1)
|
||||
: normalized;
|
||||
if (!bare) return 'other';
|
||||
|
||||
if (bare.startsWith('claude-')) return 'claude';
|
||||
if (bare.startsWith('gemma-')) return 'gemma';
|
||||
|
||||
// QwQ is Qwen's dedicated reasoning family — check before generic qwen prefix.
|
||||
if (bare.startsWith('qwq-') || bare === 'qwq') return 'qwen-reasoning';
|
||||
|
||||
// Qwen3.6-A3B series unified thinking+base into one SKU with default-on
|
||||
// reasoning (per HF model card + LOCKED 2026-04-19 target-model decision).
|
||||
// The slug lacks explicit reasoning markers, so special-case the A3B
|
||||
// variants of the 3.6 family before the generic qwen prefix check.
|
||||
if (bare.startsWith('qwen3.6-') && bare.includes('a3b')) return 'qwen-reasoning';
|
||||
|
||||
if (bare.startsWith('qwen')) {
|
||||
const hasReasoningMarker = QWEN_REASONING_MARKERS.some(marker => bare.includes(marker));
|
||||
return hasReasoningMarker ? 'qwen-reasoning' : 'qwen-instruction';
|
||||
}
|
||||
|
||||
if (bare.startsWith('llama-3') || bare.startsWith('llama-4')) return 'llama';
|
||||
|
||||
return 'other';
|
||||
}
|
||||
Reference in New Issue
Block a user