zenlm/zen-router
1
0
Fork 0
mirror of https://github.com/zenlm/zen-router.git synced 2026-10-30 14:57:44 +00:00
Code Issues Packages Projects Releases Wiki Activity
Files
main
zen-router/training/catalog.yaml
T
zeekay 7f466d8fc9 zen-router: scaled + privacy-safe run on local+agentic corpus 
- catalog.yaml: 28 routable models (add claude-opus-4-5/4-1, sonnet-4-5,
  minimax-m2 from the agentic corpus; all gateway-priced). Route head 24->28.
- session_common.py: shared session-extraction (DRY) -- task heuristic, catalog
  mapping with dated-id normalization, user/assistant pairing.
- extract_agentic_dataset.py: stream zen-agentic-dataset-private, pull embedded
  Claude transcripts out of assistant turns, stratified cap. 9.9B-token corpus
  yields ~4.4k real routing rows (rest is synthetic identity SFT + git history).
- extract_local_sessions.py: refactored onto session_common.
- sft.py: --freeze-backbone -> precompute pooled embeddings once, train the 3
  linear heads only; publish heads-only (1.2MB) so private data cannot be
  memorized. Backbone stays public zen-nano-0.6b.
- eval_routing.py: strict=False load (heads-only), device-aware, latency + dist.
- README Measured: Run A (local, full) vs Run B (combined, frozen, published)
  side by side. Run B route_acc 0.791 = +25.8pts over majority, 5/28 models
  emitted -- genuine multi-model discrimination; task_acc 0.723. All caveats.

HF: https://huggingface.co/zenlm/zen-router (heads-only weights + README).
2026-07-07 17:44:37 -07:00

63 lines
5.7 KiB
YAML
Raw Permalink Blame History

# Model catalog for the route head -- the routable universe.
#
# Route-to-anything: adding a model as a routing target is ONE row here (the
# route head sizes itself from this file via `classes_from: catalog` in
# config.yaml) PLUS the serving side already dispatching it -- the Hanzo gateway
# ModelRoute for cloud providers, hanzo-engine for local. A brand-new row's
# logit is dead until training data for it exists (mask or retrain).
#
# cost_per_1k: blended USD per 1k tokens = (input+output)/2 per-million / 1000.
# Prices are the gateway's live pricing, sourced from
# /Users/a/work/hanzo/hanzoai/ai/conf/models.yaml
# (which refreshes from pricing.hanzo.ai in live_mode). The per-million
# (input/output) each row was derived from is noted inline. Rows marked
# `# verify` are current top-end models not yet in that config; their price is
# taken from the provider's public list and must be verified against
# pricing.hanzo.ai.
models:
# -- local zen (on-device via hanzo-engine, no per-token cost) --------------
- {id: zen-nano-0.6b, backend: local, provider: engine, tasks: [cheap_chat, general], cost_per_1k: 0.0} # 0.6B, runs on laptop/phone
- {id: zen-agent-4b, backend: local, provider: engine, tasks: [general, reasoning], cost_per_1k: 0.0} # 4B
- {id: zen-coder-24b, backend: local, provider: engine, tasks: [code, reasoning], cost_per_1k: 0.0} # 24B
# -- zen cloud tiers (gateway do-ai) ---------------------------------------
- {id: zen5-flash, backend: cloud, provider: do-ai, tasks: [cheap_chat, general], cost_per_1k: 0.0002} # 0.1/0.3
- {id: zen5, backend: cloud, provider: do-ai, tasks: [general, reasoning], cost_per_1k: 0.0063} # 3.0/9.6
- {id: zen5-coder, backend: cloud, provider: do-ai, tasks: [code], cost_per_1k: 0.0063} # 3.0/9.6
- {id: zen5-max, backend: cloud, provider: do-ai, tasks: [reasoning, math, long_context], cost_per_1k: 0.024} # 12.0/36.0
- {id: zen-omni, backend: cloud, provider: do-ai, tasks: [vision, general], cost_per_1k: 0.0042} # 1.8/6.6
- {id: zen-vl, backend: cloud, provider: do-ai, tasks: [vision], cost_per_1k: 0.00113} # 0.45/1.8
# -- Anthropic Claude (gateway anthropic; dated ids normalized to undated) --
- {id: claude-opus-4-8, backend: cloud, provider: anthropic, tasks: [code, reasoning, math, long_context], cost_per_1k: 0.045} # 15/75 # verify: newer than gateway opus-4-6
- {id: claude-opus-4-5, backend: cloud, provider: anthropic, tasks: [code, reasoning, math, long_context], cost_per_1k: 0.045} # 15/75 (in gateway)
- {id: claude-opus-4-1, backend: cloud, provider: anthropic, tasks: [reasoning, math, long_context], cost_per_1k: 0.045} # 15/75 (gateway claude-4-1-opus)
- {id: claude-fable-5, backend: cloud, provider: anthropic, tasks: [reasoning, creative, general], cost_per_1k: 0.045} # 15/75 # verify: GA top model, provider list
- {id: claude-sonnet-4-6, backend: cloud, provider: anthropic, tasks: [code, reasoning, creative, general], cost_per_1k: 0.009} # 3/15 (in gateway)
- {id: claude-sonnet-4-5, backend: cloud, provider: anthropic, tasks: [code, reasoning, creative, general], cost_per_1k: 0.009} # 3/15 (in gateway)
- {id: claude-haiku-4-5, backend: cloud, provider: anthropic, tasks: [cheap_chat, general], cost_per_1k: 0.003} # 1/5 (in gateway)
# -- OpenAI (gateway do-ai / openai-direct) --------------------------------
- {id: gpt-5.5, backend: cloud, provider: openai, tasks: [reasoning, code, general], cost_per_1k: 0.01} # 5/15 # verify: newer than gpt-5.2
- {id: gpt-5.4-mini, backend: cloud, provider: openai, tasks: [cheap_chat, general], cost_per_1k: 0.00313} # 1.25/5 # verify: mini tier
- {id: gpt-5.2-pro, backend: cloud, provider: openai, tasks: [reasoning, math, long_context], cost_per_1k: 0.02} # 10/30 (in gateway)
- {id: gpt-5.1-codex-max, backend: cloud, provider: openai, tasks: [code], cost_per_1k: 0.0075} # 3/12 (in gateway)
- {id: o3, backend: cloud, provider: openai, tasks: [reasoning, math], cost_per_1k: 0.025} # 10/40 (in gateway)
# -- Google Gemini (not in gateway; provider public list) ------------------
- {id: gemini-3-pro, backend: cloud, provider: google, tasks: [reasoning, long_context, vision], cost_per_1k: 0.00875} # 2.5/15 # verify
- {id: gemini-3-flash, backend: cloud, provider: google, tasks: [cheap_chat, general, long_context], cost_per_1k: 0.0014} # 0.3/2.5 # verify
# -- xAI Grok (not in gateway; provider public list) -----------------------
- {id: grok-4, backend: cloud, provider: xai, tasks: [reasoning, code, general], cost_per_1k: 0.009} # 3/15 # verify
# -- DeepSeek (gateway do-ai) ----------------------------------------------
- {id: deepseek-v3.2, backend: cloud, provider: do-ai, tasks: [code, general], cost_per_1k: 0.00069} # 0.27/1.1
- {id: deepseek-reasoner, backend: cloud, provider: do-ai, tasks: [reasoning, math], cost_per_1k: 0.0055} # 3/8
# -- Moonshot Kimi (gateway do-ai) -----------------------------------------
- {id: kimi-k2, backend: cloud, provider: do-ai, tasks: [reasoning, code, long_context], cost_per_1k: 0.005} # 2/8
# -- MiniMax (gateway fireworks) -------------------------------------------
- {id: minimax-m2, backend: cloud, provider: fireworks, tasks: [code, general], cost_per_1k: 0.0025} # 1/4 (gateway minimax-m2p5)
Reference in New Issue View Git Blame Copy Permalink