protein-model-registry / selection_policy.yaml
theoschiff-biie's picture
Update model selection policy
4f6f2bf verified
Raw
History Blame Contribute Delete
4.36 kB
---
# Model Selection Policy
# Used by model-selection/SKILL.md as the routing and ranking source of truth.
# Edit this file to adjust priorities without touching the skill logic.
default_output_count: 3
always_include_budget_option: true
always_include_frontier_option: true
# Scoring weights (must sum to 1.0)
weights:
task_fit: 0.40 # primary_tasks match
modality_fit: 0.20 # input/output/conditioning modality match
compute_fit: 0.15 # compute_tier within stated budget
openness_fit: 0.10 # open_weights + commercial_use alignment
maturity_score: 0.10 # established > recent > experimental
deployment_practicality: 0.05 # single_gpu_feasible, no gating, etc.
# Penalties applied on top of the weighted score
penalties:
gated: 0.20 # gated=true and user wants frictionless
over_budget_compute: 0.25 # compute_tier exceeds stated budget
unknown_commercial_status: 0.15 # commercial_use=unknown when deployment is commercial
legacy_model: 0.10 # status=legacy or exclusion_tags contains prefer_*_over_this
non_commercial_commercial_deploy: 0.30 # commercial_use=non_commercial_only|research_only when commercial
# Default assumed compute budget when none stated
default_compute_budget: single_gpu_midmem
# Task → ordered candidate list (first is preferred; apply compute filter after)
routing:
embedding:
- esmc_300m # strongest small embedder (non-commercial)
- esm2_t33_650m # best commercial default
- saprot_650m # structure-aware, if structure available
- ankh3_large # multi-task pretrain, strong generalization
- prot_t5_xl_enc # T5 ecosystem; slower but competitive
- esm2_t30_150m # budget fallback
variant_effect_prediction:
- esm1v_650m # purpose-built masked-marginal scorer
- tranception # autoregressive + retrieval; best on fitness landscapes
- saprot_650m # structure-aware variant effect if structure available
- esm2_t33_650m # general fallback with PLL scoring
- esm2_t6_8m # budget option with PLL scoring
fitness_prediction:
- tranception
- esm1v_650m
- esm2_t33_650m
sequence_generation:
- protgpt2 # established decoder-only baseline
- dplm_650m # diffusion; better diversity
- esm3_sm_open # generative multimodal (non-commercial)
controlled_generation:
- esm3_sm_open # best open for structure/function conditioned
- dplm2_150m # joint seq+structure, lighter
- nv_la_proteina # frontier all-atom (gated, research-only)
folding:
- esmfold_v1 # fast MSA-free, open weights
- boltz_2 # better for complexes / affinity
- dplm2_150m # lighter; co-designs sequence + structure
inverse_folding:
- esm_if1 # purpose-built, small, commercial
- dplm_650m # diffusion-based; good diversity
- dplm2_150m # joint seq+structure design
motif_scaffolding:
- dplm_650m
- dplm2_150m
- nv_la_proteina # frontier all-atom scaffolding
binder_design:
- boltzgen_1 # purpose-built binder generator
- nv_la_proteina # all-atom frontier design
- boltz_2 # complex prediction + affinity
complex_prediction:
- boltz_2
- boltzgen_1
binding_affinity_prediction:
- boltz_2
nucleotide_representation:
- nucleotide_transformer_v2_500m
biomedical_nlp:
- biomedbert_base # MIT, PubMed abstracts
- biobert_base # Apache 2.0, PubMed + PMC
# Hard filter rules (applied before scoring)
# Models failing any applicable rule are excluded from candidates.
hard_filters:
- rule: task_mismatch
description: Exclude if task not in model primary_tasks or secondary_tasks
- rule: compute_over_budget
description: Exclude if compute_tier > stated budget (unless no candidate passes)
- rule: gated_frictionless
description: Exclude gated models when user explicitly requests frictionless access
- rule: commercial_restriction
description: Exclude non_commercial_only/research_only models when commercial deployment is implied
- rule: wrong_category
description: Exclude nucleotide_language_model and biomedical_text_model for protein tasks (and vice versa)