Navigating These Docs

These docs are organized by workflow: start with Weaver basics, then move into the API surface, and finally connect NexRL and NexAU for agentic model training.

Documentation Structure

Getting Started

For first-time Weaver users:

• What is Weaver?: Weaver's purpose, abstractions, and use cases.
• Installation: install the Python SDK, configure an API key, and run a first training example.

Weaver API

For writing training scripts or integrating an existing framework:

• Training and Sampling: ServiceClient, TrainingClient, Datum, training loops, sampling, and logprobs.
• Loss Functions: cross entropy, importance sampling, PPO/GRPO-related losses, and custom loss paths.
• Saving and Loading: sampler exports, durable checkpoints, optimizer restore, and TTL.
• Model Lineup: supported models, training modes, long-context naming, and selection guidance.

Agentic Model Training

For training models that use tools, run multi-step tasks, or learn from agent traces:

• Finetune Models: use NexRL to orchestrate rollout, trajectory pools, rewards, and Weaver training.
• Build & Finetune Agents: define NexAU agents and connect their trajectories to NexRL/Weaver.

Suggested Paths

If you want the smallest working example:

1. Installation
2. Training and Sampling
3. Saving and Loading

If you are doing RL or agent training:

1. Loss Functions
2. Finetune Models
3. Build & Finetune Agents

If you are integrating experiment management, pay special attention to:

• Checkpoint save strategy and TTL.
• The default 1-hour TTL for sampler weight exports.
• How performance_tier affects throughput and cost.
• Async OperationHandle results from wait=False.