mirror of
https://github.com/zenlm/zen-router.git
synced 2026-10-30 14:57:44 +00:00
zeekay
f888212afb
Reads a prompt, picks the best model. 0.6B encoder-pooled router with task/route/feature heads; SFT from eval profiles + usage ledger, GRPO on quality - lambda*cost - mu*latency; serves via hanzo-engine and plugs the hanzo-router Classifier and enso Featurizer seams. Apache-2.0.
27 lines
703 B
YAML
27 lines
703 B
YAML
# zen-router training config
|
|
base_model: zenlm/zen-nano-0.6b
|
|
output_dir: out/zen-router
|
|
|
|
tasks: [code, reasoning, math, creative, vision, long_context, cheap_chat, general]
|
|
|
|
heads:
|
|
task: {type: classification, classes: 8}
|
|
route: {type: classification, classes_from: catalog} # model catalog ids
|
|
features: {type: projection, dim: 256} # bandit feature head
|
|
|
|
catalog: training/catalog.yaml # ModelCard ids + cost_per_1k, mirrors router registry
|
|
|
|
sft:
|
|
lr: 2.0e-5
|
|
epochs: 2
|
|
batch_size: 64
|
|
max_seq_len: 2048
|
|
pooled: last_token
|
|
|
|
grpo:
|
|
# reward = quality - lambda*cost - mu*latency (matches serving SLO)
|
|
lambda_cost: 0.35
|
|
mu_latency: 0.15
|
|
group_size: 8
|
|
lr: 5.0e-6
|