> ## Documentation Index > Fetch the complete documentation index at: https://docs.openclaas.com/llms.txt > Use this file to discover all available pages before exploring further. # Configuration Reference > Complete Hydra configs and environment variables for all CLaaS backends and eval # Configuration Reference Exhaustive reference for every Hydra config field and environment variable across all CLaaS backends and the eval harness. For getting-started guides, see the individual backend pages. ## Local Backend See [Local GPU Backend](/training/local) for setup instructions. ### Hydra Config (`claas/core/configs/local.yaml`) ```yaml theme={null} defaults: - _core_local_schema - _self_ mode: local feedback_log_dir: ./data/feedback lora_root: /loras storage_backend: local_fs allowed_init_base_models: - Qwen/Qwen3-8B vllm_base_url: http://127.0.0.1:8000 feedback_lock_timeout_s: 120.0 feedback_wake_on_failure: true feedback_min_free_vram_gb: 20.0 feedback_sleep_verify_timeout_s: 30.0 feedback_drain_timeout_s: 30.0 base_model_id: Qwen/Qwen3-8B attn_implementation: flash_attention_2 completion_cache_size: 100 ``` ### Environment Variables These correspond to the `LocalConfig` dataclass fields and Docker Compose `.env` variables. | Variable | Default | Description | | --------------------------------- | ----------------------- | ------------------------------------------------------------- | | `TELEGRAM_BOT_TOKEN` | *(required)* | Bot token from [@BotFather](https://t.me/BotFather) | | `HF_TOKEN` | — | HuggingFace token for gated models | | `MODEL` | `Qwen/Qwen3-8B` | Base model ID | | `MAX_MODEL_LEN` | `32768` | Max sequence length for vLLM | | `GPU_MEMORY_UTILIZATION` | `0.70` | GPU VRAM fraction for vLLM | | `LORA_NAME` | `openclaw/assistant` | LoRA adapter identity | | `CLAAS_API_PORT` | `8080` | Host port for CLaaS API | | `OPENCLAW_PORT` | `18789` | Host port for OpenClaw gateway | | `CLAAS_DISTILL_EXECUTION_MODE` | `local` | Training engine | | `CLAAS_BASE_MODEL_ID` | `Qwen/Qwen3-8B` | Base model for LoRA training | | `CLAAS_LORA_ROOT` | `/loras` | Root directory for LoRA adapter storage | | `CLAAS_STORAGE_BACKEND` | `local_fs` | Storage backend: `local_fs`, `modal_volume`, or `tinker_json` | | `CLAAS_ALLOWED_INIT_BASE_MODELS` | `Qwen/Qwen3-8B` | Comma-separated allowlist for `/v1/lora/init` | | `CLAAS_COMPLETION_CACHE_SIZE` | `100` | Completion cache entries to retain | | `FEEDBACK_LOG_DIR` | `./data/feedback` | Directory for structured feedback JSON logs | | `VLLM_BASE_URL` | `http://127.0.0.1:8000` | vLLM server URL | | `VLLM_API_KEY` | `sk-local` | API key for vLLM authentication | | `CLAAS_ATTN_IMPLEMENTATION` | `flash_attention_2` | Attention backend (`sdpa`, `flash_attention_2`) | | `FEEDBACK_LOCK_TIMEOUT_S` | `120` | Per-LoRA lock timeout in seconds | | `FEEDBACK_WAKE_ON_FAILURE` | `true` | Wake vLLM if the distill step fails | | `FEEDBACK_MIN_FREE_VRAM_GB` | `20` | Minimum free VRAM (GB) before training starts | | `FEEDBACK_SLEEP_VERIFY_TIMEOUT_S` | `30` | Timeout (s) waiting for vLLM to sleep | | `FEEDBACK_DRAIN_TIMEOUT_S` | `30` | Timeout (s) draining vLLM queue before sleep | *** ## Tinker Backend See [Tinker SDK Backend](/training/tinker) for setup instructions. ### Hydra Config (`claas/core/configs/tinker.yaml`) ```yaml theme={null} defaults: - _core_tinker_schema - _self_ mode: tinker feedback_log_dir: ./data/feedback lora_root: /loras storage_backend: local_fs allowed_init_base_models: - Qwen/Qwen3-8B vllm_base_url: http://127.0.0.1:8000 tinker_base_model: Qwen/Qwen3-30B-A3B tinker_state_path: /data/tinker_state.json completion_cache_size: 100 ``` ### Environment Variables These correspond to the `TinkerConfig` dataclass fields and Docker Compose `.env.tinker` variables. | Variable | Default | Description | | -------------------------------- | ---------------------------- | ---------------------------------------------------- | | `TELEGRAM_BOT_TOKEN` | *(required)* | Bot token from [@BotFather](https://t.me/BotFather) | | `TINKER_API_KEY` | *(required)* | API key for Tinker SDK | | `MODEL` | `Qwen/Qwen3-30B-A3B` | Base model ID (Tinker name) | | `LORA_NAME` | `openclaw/assistant` | LoRA adapter identity | | `CLAAS_API_PORT` | `8080` | Host port for CLaaS API | | `OPENCLAW_PORT` | `18789` | Host port for OpenClaw gateway | | `TINKER_PROXY_PORT` | `8000` | Host port for Tinker inference proxy | | `FEEDBACK_BATCH_SIZE` | `4` | Samples per feedback batch before triggering distill | | `CLAAS_DISTILL_EXECUTION_MODE` | `tinker` | Training engine | | `CLAAS_LORA_ROOT` | `/loras` | Root directory for LoRA adapter storage | | `CLAAS_STORAGE_BACKEND` | `local_fs` | Storage backend (use `tinker_json` for Tinker) | | `CLAAS_ALLOWED_INIT_BASE_MODELS` | `Qwen/Qwen3-8B` | Comma-separated allowlist for `/v1/lora/init` | | `CLAAS_COMPLETION_CACHE_SIZE` | `100` | Completion cache entries to retain | | `FEEDBACK_LOG_DIR` | `./data/feedback` | Directory for structured feedback JSON logs | | `CLAAS_TINKER_API_KEY` | *(required)* | Tinker SDK API key | | `CLAAS_TINKER_BASE_MODEL` | `Qwen/Qwen3-30B-A3B` | Hosted model for distillation | | `CLAAS_TINKER_STATE_PATH` | `~/.claas/tinker_state.json` | Local path for Tinker LoRA state file | | `VLLM_BASE_URL` | `http://127.0.0.1:8000` | URL of the Tinker inference proxy | | `VLLM_API_KEY` | `sk-local` | API key for inference proxy | *** ## Modal Backend See [Modal Cloud Backend](/training/modal) for setup instructions. ### Hydra Config (`claas/core/configs/modal.yaml`) ```yaml theme={null} defaults: - _core_modal_schema - _self_ mode: modal feedback_log_dir: ./data/feedback lora_root: /loras storage_backend: modal_volume allowed_init_base_models: - Qwen/Qwen3-8B vllm_base_url: http://127.0.0.1:8000 feedback_lock_timeout_s: 120.0 feedback_wake_on_failure: true feedback_min_free_vram_gb: 20.0 feedback_sleep_verify_timeout_s: 30.0 feedback_drain_timeout_s: 30.0 completion_cache_size: 100 ``` ### Environment Variables These correspond to the `ModalConfig` dataclass fields. | Variable | Default | Description | | --------------------------------- | ----------------------- | ---------------------------------------------- | | `HF_TOKEN` | — | HuggingFace token for gated models | | `CLAAS_DISTILL_EXECUTION_MODE` | `modal` | Training engine | | `CLAAS_BASE_MODEL_ID` | `Qwen/Qwen3-8B` | Base model for LoRA training | | `CLAAS_LORA_ROOT` | `/loras` | Root directory for LoRA adapter storage | | `CLAAS_STORAGE_BACKEND` | `modal_volume` | Storage backend (use `modal_volume`) | | `CLAAS_ALLOWED_INIT_BASE_MODELS` | `Qwen/Qwen3-8B` | Comma-separated allowlist for `/v1/lora/init` | | `CLAAS_COMPLETION_CACHE_SIZE` | `100` | Completion cache entries to retain | | `FEEDBACK_LOG_DIR` | `./data/feedback` | Directory for structured feedback JSON logs | | `VLLM_BASE_URL` | `http://127.0.0.1:8000` | vLLM server URL (within Modal) | | `VLLM_API_KEY` | `sk-local` | API key for vLLM authentication | | `FEEDBACK_LOCK_TIMEOUT_S` | `120` | Per-LoRA lock timeout in seconds | | `FEEDBACK_WAKE_ON_FAILURE` | `true` | Wake vLLM if the distill step fails | | `FEEDBACK_MIN_FREE_VRAM_GB` | `20` | Minimum free VRAM (GB) before training starts | | `FEEDBACK_SLEEP_VERIFY_TIMEOUT_S` | `30` | Timeout (s) waiting for vLLM to sleep | | `FEEDBACK_DRAIN_TIMEOUT_S` | `30` | Timeout (s) draining vLLM queue before sleep | | `CLAAS_HF_SECRET_NAME` | — | Name of Modal Secret containing HF credentials | *** ## Eval Harness See [Configuration & Running](/eval/running) for setup instructions. ### Hydra Config (`claas/eval/configs/base.yaml`) ```yaml theme={null} mode: tinker claas_url: http://localhost:8080 base_model: Qwen/Qwen3-30B-A3B preferences: - no_emoji - concise - identity metrics: - logprob - compliance - general - collapse collapse_steps: [0, 5, 10, 15, 19] plots: true num_steps: 20 batch_size: 4 steps_per_batch: 1 feedback_repetitions: 1 seed: 42 lora_id_prefix: eval output_dir: ./data/evals openclaw_url: http://localhost:18789 ``` ### Config Fields | Field | Type | Default | Description | | ---------------------- | ------------- | ------------------------------------------ | --------------------------------------------------------- | | `mode` | `str` | `tinker` | Execution backend: `local`, `tinker`, or `modal` | | `claas_url` | `str` | `http://localhost:8080` | CLaaS API endpoint | | `base_model` | `str` | `Qwen/Qwen3-30B-A3B` | Base model for LoRA init (use Tinker name in Tinker mode) | | `preferences` | `list[str]` | `[no_emoji, concise, identity]` | Preferences to train and evaluate | | `metrics` | `list[str]` | `[logprob, compliance, general, collapse]` | Metrics to evaluate per step | | `num_steps` | `int` | `20` | Number of feedback steps per preference | | `batch_size` | `int` | `4` | Samples per feedback batch | | `steps_per_batch` | `int` | `1` | Gradient updates per batch | | `feedback_repetitions` | `int` | `1` | Times to repeat feedback string per sample | | `collapse_steps` | `list[int]` | `[0, 5, 10, 15, 19]` | Steps where collapse metric runs | | `plots` | `bool` | `true` | Generate matplotlib plots | | `seed` | `int` | `42` | Random seed for reproducibility | | `lora_id_prefix` | `str` | `eval` | Prefix for eval LoRA adapter IDs | | `output_dir` | `str` | `./data/evals` | Directory for eval output | | `openclaw_url` | `str \| null` | `http://localhost:18789` | OpenClaw gateway URL (null = use CLaaS proxy directly) | ### Environment Variables (Secrets) Secrets are resolved from environment variables at runtime, not stored in config files. | Variable | Required for | Purpose | | -------------------------------- | ---------------- | -------------------------------------------- | | `CLAAS_TINKER_API_KEY` | Tinker mode | Tinker SDK authentication | | `CLAAS_TINKER_BASE_MODEL` | Tinker mode | Must match `base_model` in config | | `CLAAS_DISTILL_EXECUTION_MODE` | API server | `tinker`, `local`, or `modal` | | `CLAAS_ALLOWED_INIT_BASE_MODELS` | API server | Comma-separated allowed models for LoRA init | | `VLLM_API_KEY` | Local mode | vLLM server auth token | | `OPENCLAW_GATEWAY_TOKEN` | OpenClaw mode | Auth token for OpenClaw gateway | | `GEMINI_API_KEY` | `general` metric | Gemini-based capability evaluation |