🚀 Project website | 📚 Docs | 💬 Discord

CLaaS: Continual Learning as a Service

Continual learning as-a-service (CLaaS): distill context into weights to retain knowledge and make your tokens count.

  

Installation

Prerequisites: Python 3.11+, uv, and Docker.

CLaaS supports three execution engines. Pick the one that matches your setup:

Engine	GPU required	Status
Local	Yes (>= 24 GB VRAM)	Available
Tinker (recommended)	No	Available
Modal	No (remote GPU)	Coming soon

Local (GPU)

Requires an NVIDIA GPU with >= 24 GB VRAM (L40S, A100, RTX 4090, RTX 5090, etc.) and the NVIDIA Container Toolkit. Local GPU setup is pinned to CUDA 12.8 (cu128 / NGC PyTorch 25.01), so your NVIDIA driver must support CUDA 12.8. The native (non-Docker) setup is tested on an RTX 5090.

Docker Compose (recommended):

cd docker
cp .env.local.example .env
# Edit .env — set TELEGRAM_BOT_TOKEN (from @BotFather)
docker compose --profile local up --build

This brings up vLLM with Qwen3-8B, the CLaaS feedback API, and OpenClaw's Telegram gateway. See docker/README.md for details.

Manual install:

uv sync --extra local

Then start vLLM and the API yourself. See Quick Start and docker/README.md for the full supervised local stack.

If you use Claude Code, /setup-local <TELEGRAM_BOT_TOKEN> installs all deps and starts the full local stack automatically.

Tinker (no GPU)

Uses the Tinker SDK for hosted distillation and inference. Requires a TINKER_API_KEY.

Docker Compose (recommended):

cd docker
cp .env.tinker.example .env.tinker
# Edit .env.tinker — set TELEGRAM_BOT_TOKEN and TINKER_API_KEY
docker compose --env-file .env.tinker --profile tinker up --build

Manual install:

uv sync --extra tinker
uv pip install --python .venv/bin/python --index-url https://download.pytorch.org/whl/cpu torch

If you use Claude Code, /setup-tinker deploys the Tinker Docker stack. Use /clear-tinker-storage to delete all Tinker checkpoints and free storage.

Modal (remote GPU) — Coming Soon

Runs distillation remotely on Modal (L40S). Requires a Modal account.

uv sync --extra local
uv run modal token new

Deploy:

# Set HF_TOKEN if using gated models
export HF_TOKEN=...
export CLAAS_BASE_MODEL_ID=Qwen/Qwen3-8B
uv run modal deploy -m claas.modal.deploy

The deployed app exposes the same API at https://your-app--claas-distill-fastapi-app.modal.run. LoRAs are stored in the claas-loras Modal Volume.

If you use Claude Code, /setup-modal deploys the CLaaS distillation service to Modal.

Quick Start

For manual (non-Docker) local setup:

# 1. Start vLLM with LoRA support
vllm serve Qwen/Qwen3-8B --host 0.0.0.0 --port 8000 \
  --enable-lora --lora-modules my-lora=/loras/user/my-lora-init

# 2. Start the CLaaS API
uv run python -m claas.api --config-name local

# 3. Initialize a LoRA adapter
curl -X POST http://localhost:8080/v1/lora/init \
  -H "Content-Type: application/json" \
  -d '{"lora_id": "user/my-lora"}'

# 4. Send a feedback update
curl -X POST http://localhost:8080/v1/feedback \
  -H "Content-Type: application/json" \
  -d '{
    "requests": [{
      "lora_id": "user/my-lora",
      "prompt": "Write a function to calculate factorial",
      "response": "def factorial(n): ...",
      "feedback": "Good recursive solution"
    }]
  }'

For the full supervised local stack (vLLM + gateway + Telegram), see docker/README.md.

Hybrid engine

The locally hosted request path is driven by a hybrid engine that switches between:

• Serving mode: route request traffic through vLLM (local or remote) for low-latency generation.
• Update mode: run a single self-distillation LoRA step using the provided feedback to adapt the adapter.

In practice, the flow is: request is answered by vLLM, then the engine performs (or schedules) the training step, and subsequent requests can use the updated adapter. The engine can prefer local or remote teacher inference depending on teacher_mode.

Eval Harness

The eval harness runs automated feedback loops against a live CLaaS stack and measures whether training shifts the model toward preferred behaviours without collapsing. Configuration uses Hydra with YAML configs.

# Install tinker + dev dependencies
uv sync --extra tinker --extra dev

# Run conciseness eval for 20 steps (Tinker mode, no GPU)
CLAAS_TINKER_API_KEY="tml-..." \
  uv run python -m claas.eval 'preferences=[concise]' num_steps=20

Override any config field via Hydra's key=value syntax. The default config is in claas/eval/configs/base.yaml. See claas/eval/README.md for full documentation including metrics, config reference, setup steps, and known gotchas.

Important: When using Tinker mode, base_model must use Tinker's model name (e.g. Qwen/Qwen3-30B-A3B), not the HuggingFace name (Qwen/Qwen3-Coder-30B-A3B-Instruct). Tinker's sampling API accepts either, but the LoRA training API only accepts the Tinker name.

Dashboard

The CLaaS API serves a built-in dashboard at /v1/dashboard showing recent feedback batches, training metrics, and timing breakdowns. Each row is a batch — expand it to see individual samples and detailed metrics.

The eval dashboard at /v1/eval displays results from running the eval harness against your model.

References

1. Kleine Buening et al. (2026). "Aligning Language Models from User Interactions." Paper · Code · Hindsight template
2. Hübotter et al. (2026). "Reinforcement Learning via Self-Distillation." arXiv:2601.20802
3. SDPO Reference Implementation: https://github.com/lasgroup/SDPO
4. Modal GPU Memory Snapshots: https://modal.com/blog/gpu-mem-snapshots
5. vLLM: https://github.com/vllm-project/vllm
6. PEFT/LoRA: https://github.com/huggingface/peft
7. Tinker SDPO training reference (continualcode): https://github.com/sdan/continualcode

License

MIT