SitongGong's picture
merge code dirs 1031 (part 4)
ff8b94c verified
|
Raw History Blame Contribute Delete
4.3 kB

The Training Loop

SkillOpt's core insight: optimizing natural-language skill documents follows the same structure as training neural networks.

Overview

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚                    Training Loop                         β”‚
β”‚                                                         β”‚
β”‚  for epoch in epochs:                                   β”‚
β”‚    for step in steps:                                   β”‚
β”‚      1. Rollout   β€” Target executes tasks              β”‚
β”‚      2. Reflect   β€” Optimizer analyzes trajectories       β”‚
β”‚      3. Aggregate β€” Hierarchical merge of patches       β”‚
β”‚      4. Select    β€” Rank & clip edits (learning rate)   β”‚
β”‚      5. Update    β€” Apply patches to skill doc          β”‚
β”‚      6. Gate      β€” Validate & accept/reject            β”‚
β”‚                                                         β”‚
β”‚    Epoch Boundary:                                       β”‚
β”‚      β€’ Slow Update (longitudinal comparison & guidance) β”‚
β”‚      β€’ Meta Skill  (cross-epoch strategy memory)        β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

Stage Details

1. Rollout (Forward Pass)

The target model executes tasks using the current skill document as its prompt. Each task produces a trajectory and a score.

# Analogy: forward pass through the network
predictions = model(input, skill_document)
scores = evaluate(predictions, ground_truth)

2. Reflect (Backward Pass)

The optimizer model analyzes trajectory minibatches and produces edit patches β€” structured suggestions for improving the skill document. Failure minibatches are always eligible for analysis; successful trajectories are also analyzed unless gradient.failure_only is enabled. Independent minibatches can run concurrently according to gradient.analyst_workers.

# Analogy: computing gradients
gradients = loss.backward()  # β†’ edit patches

3. Aggregate

Semantically similar edit patches are merged to avoid redundant edits.

4. Select (Gradient Clipping)

Edits are ranked by relevance score. The learning_rate parameter caps how many edits are applied per step β€” just like gradient clipping prevents overshooting.

# Analogy: gradient clipping + optimizer step size
selected = top_k(edits, k=learning_rate)

The lr_scheduler adjusts this over training:

  • cosine: Start aggressive, taper smoothly
  • linear: Linear decay
  • constant: Fixed rate

5. Update (Parameter Update)

Selected edits are applied to the skill document, producing a new version.

6. Gate (Validation)

The updated skill is evaluated on a selection split (analogous to a validation set). With the gate enabled, the candidate is accepted only when its configured gate score (hard, soft, or mixed) is strictly higher than the current skill's score. With evaluation.use_gate: false, validation is still recorded but candidates are force-accepted.

Epoch Boundary Mechanisms

Slow Update

At the end of each epoch (starting from epoch 2), the system performs a longitudinal comparison: it rolls out both the previous epoch's skill and the current skill on the same samples, categorizes items as improved/regressed/persistent-fail/stable-success, then generates high-level guidance for the skill document. Depending on optimizer.slow_update_gate_with_selection, that guidance is either checked on the selection split or applied unconditionally. Its purpose is to counter cross-epoch forgetting.

Meta Skill

A meta-skill memory accumulates high-level strategy notes across the training run. Starting at the end of epoch 2, the optimizer compares the previous and current epoch, writes a compact memory, and provides the prior epoch's memory as additional context during later reflection and update stages.

Next Steps