Phase 6: Future capabilities - LLM config generation, HyperNetworkCombiner, Nash-MTL

[w4nderlust]

Summary

Phase 6 of the Ludwig modernization: advanced capabilities that extend Ludwig's competitive positioning.

1. LLM-Driven Config Generation

Describe your ML task in natural language and get a valid Ludwig config:

from ludwig.config_generation import generate_config

config = generate_config(
    'I have customer data with age, income, and purchase history. '
    'I want to predict churn (binary) and lifetime value (number).'
)
# Returns a validated Ludwig config dict

Uses Ludwig's JSON Schema as context for the LLM. Supports both Anthropic and OpenAI APIs. Generated configs are validated against Ludwig's Pydantic schema before returning.

2. HyperNetworkCombiner

A hypernetwork generates the weights of processing layers based on a conditioning modality. Instead of combining features additively, metadata fundamentally changes how other features are processed.

combiner:
  type: hypernetwork
  hidden_size: 128
  hyper_hidden_size: 64
  output_size: 128

Based on HyperFusion (arXiv 2403.13319, 2024). Unique Ludwig differentiator not available in other frameworks.

3. Nash-MTL Loss Balancer

Nash bargaining for multi-task weight allocation (Navon et al., ICML 2022). Finds the cooperative game-theoretic solution where weights are inversely proportional to task losses, equalizing marginal contributions.

trainer:
  loss_balancing: nash_mtl

For power users with many conflicting output features. More principled than heuristic methods. For most use cases, FAMO or uncertainty weighting are sufficient.

Test plan

• 1223 existing tests pass (0 regressions, 1 pre-existing flaky dask test)
• Pre-commit all clean
• HyperNetworkCombiner registered and importable
• Nash-MTL registered in LOSS_BALANCER_REGISTRY
• CI