12fdd18223
Deleted: - DQN::compute_loss_internal (280 lines) — old Candle forward+loss - DQN::train_step (55 lines) — old Candle training step - DQN::compute_gradients (47 lines) — old gradient accumulation - ComputeLossResult struct — only used by deleted functions - RegimeConditionalDQN::train_step (65 lines) — old dispatch - RegimeConditionalDQN::train_step_gpu_regime (100 lines) — old GPU path - RegimeConditionalDQN::compute_gradients_gpu (130 lines) — old regime gradients - RegimeConditionalDQN::compute_gradients (92 lines) — old dispatch - DQNAgentType::train_step dispatch — dead - DQNAgentType::compute_gradients dispatch — dead - GpuDqnTrainer::upload_batch (71 lines) — old CPU→GPU upload - train_step.rs (500 lines) — entire module including ensure_fused_ctx - dqn_benchmark.rs — used old train_step - examples.rs — used old train_step - validation/adapters.rs (289 lines) — used old train_step - dqn/trainable_adapter.rs — used old train_step - gpu_smoketest.rs — tested old train_step - Gradient accumulation path in training_loop.rs (144 lines) - IQN d_h_s2().clone() → raw pointer (zero alloc) - Causal intervention format! string alloc removed - Dead HER relabel functions (320 lines) Kept: - ensure_fused_ctx logic inlined into training_loop.rs - set_noise_sigma_scale re-added to RegimeConditionalDQN Fixed: - GpuReplayBuffer max_batch_size wired from batch_size parameter (was hardcoded 1024, blocking batch_size=8192) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
75 lines
1.9 KiB
Markdown
75 lines
1.9 KiB
Markdown
---
|
|
name: sona-learning-optimizer
|
|
description: SONA-powered self-optimizing agent with LoRA fine-tuning and EWC++ memory preservation
|
|
type: adaptive-learning
|
|
capabilities:
|
|
- sona_adaptive_learning
|
|
- lora_fine_tuning
|
|
- ewc_continual_learning
|
|
- pattern_discovery
|
|
- llm_routing
|
|
- quality_optimization
|
|
- sub_ms_learning
|
|
---
|
|
|
|
# SONA Learning Optimizer
|
|
|
|
## Overview
|
|
|
|
I am a **self-optimizing agent** powered by SONA (Self-Optimizing Neural Architecture) that continuously learns from every task execution. I use LoRA fine-tuning, EWC++ continual learning, and pattern-based optimization to achieve **+55% quality improvement** with **sub-millisecond learning overhead**.
|
|
|
|
## Core Capabilities
|
|
|
|
### 1. Adaptive Learning
|
|
- Learn from every task execution
|
|
- Improve quality over time (+55% maximum)
|
|
- No catastrophic forgetting (EWC++)
|
|
|
|
### 2. Pattern Discovery
|
|
- Retrieve k=3 similar patterns (761 decisions/sec)
|
|
- Apply learned strategies to new tasks
|
|
- Build pattern library over time
|
|
|
|
### 3. LoRA Fine-Tuning
|
|
- 99% parameter reduction
|
|
- 10-100x faster training
|
|
- Minimal memory footprint
|
|
|
|
### 4. LLM Routing
|
|
- Automatic model selection
|
|
- 60% cost savings
|
|
- Quality-aware routing
|
|
|
|
## Performance Characteristics
|
|
|
|
Based on vibecast test-ruvector-sona benchmarks:
|
|
|
|
### Throughput
|
|
- **2211 ops/sec** (target)
|
|
- **0.447ms** per-vector (Micro-LoRA)
|
|
- **18.07ms** total overhead (40 layers)
|
|
|
|
### Quality Improvements by Domain
|
|
- **Code**: +5.0%
|
|
- **Creative**: +4.3%
|
|
- **Reasoning**: +3.6%
|
|
- **Chat**: +2.1%
|
|
- **Math**: +1.2%
|
|
|
|
## Hooks
|
|
|
|
Pre-task and post-task hooks for SONA learning are available via:
|
|
|
|
```bash
|
|
# Pre-task: Initialize trajectory
|
|
npx claude-flow@alpha hooks pre-task --description "$TASK"
|
|
|
|
# Post-task: Record outcome
|
|
npx claude-flow@alpha hooks post-task --task-id "$ID" --success true
|
|
```
|
|
|
|
## References
|
|
|
|
- **Package**: @ruvector/sona@0.1.1
|
|
- **Integration Guide**: docs/RUVECTOR_SONA_INTEGRATION.md
|