# Copy to configs/llm.local.yaml and edit. Used for the Phase-3/4 local
# transformers + LoRA path (LocalHFLogitsProvider). For OpenAI-compatible API
# or local OpenAI-compat servers (vLLM, Ollama, LM Studio), use llm.yaml.

provider: local_hf
model: Qwen/Qwen3-8B
# Optional: path to a LoRA adapter trained by scripts/train_lora.py
lora_adapter: null  # e.g. reports/training/v1/lora_final

# bf16 / fp16 / fp32. bf16 is the recommended default on A100.
dtype: bf16

# Set to "cuda" to put the whole model on GPU; "auto" to let HF accelerate
# device-map across two A100 cards. For 8B + LoRA + bf16 a single A100 40GB
# is enough.
device_map: auto

# First-token classification protocol. Tokens to read logits for.
# The score is softmax over (yes_token_logit, no_token_logit) at decode step 0.
yes_tokens: ["Yes", " Yes", "YES"]
no_tokens: ["No", " No", "NO"]

# How many extra new tokens after the first to record (for prompt audit only;
# scoring does not depend on them).
trace_max_new_tokens: 4

# Used by NodeTextSummarizer / MetapathTextSummarizer (Phase 2).
# The summarizer uses the SAME backbone unless summarizer_model is set.
summarizer:
  model: null              # null = reuse `model`
  b_node: 10
  b_meta: 10
  cache_dir: reports/cache/text_summary
  task_agnostic_prompt: "Summarize the text within {budget} tokens."
  max_input_tokens: 4096

# Embedder used by MarkovDiffusionTrimmer (Phase 2).
embedder:
  model: sentence-transformers/all-MiniLM-L6-v2
  device: cuda
  cache_dir: reports/cache/embeddings