config/default.yaml

# Pushback in the Wild — central configuration (plan v2.0, June 2026)
# All model ids/slugs web-verified June 2026. See references.md for sources.

random_seed: 20260512

paths:
  raw_dir: dataset/raw
  regex_filtered: dataset/raw/regex_filtered.jsonl
  classified_events: dataset/classified_events.jsonl
  confirmed_pushbacks: dataset/confirmed_pushbacks.jsonl
  pushbackchat_3k: dataset/pushbackchat_3k_v1.json
  cache_dir: results/cache
  fresh_responses_dir: results/fresh_responses
  behaviors_dir: results/behaviors
  analysis_results: analysis/results
  ground_truth: analysis/ground_truth.csv
  pushback_validation: analysis/pushback_validation.csv

dataset:
  source: allenai/WildChat-4.8M     # ODC-BY. Attribution required.
  split: train
  language_filter: en               # English only (plan)
  min_turns: 3                      # need >=3 turns so a pushback can exist in user turn 2+
  license: ODC-BY
  attribution: "Derived from allenai/WildChat-4.8M (AI2), licensed ODC-BY."

# ---- JUDGE (classification only; does NOT generate) ----
judge:
  provider: gemini
  model: gemini-3.1-flash-lite
  temperature: 0.0
  max_output_tokens: 128
  thinking_budget: 0
  rpm: 120

# ---- GENERATION TARGETS (4 OpenRouter models, architecturally diverse) ----
generation:
  temperature: 0.0
  max_tokens: 512
  rpm: 200                         # OpenRouter paid tier; tune per account
models:
  - name: deepseek_v4_flash
    provider: openrouter
    slug: deepseek/deepseek-v4-flash
    architecture: dense-moe-reasoning
  - name: llama_3_3_70b
    provider: openrouter
    slug: meta-llama/llama-3.3-70b-instruct
    architecture: dense-transformer
  - name: qwen3_235b_a22b_2507
    provider: openrouter
    slug: qwen/qwen3-235b-a22b-2507
    architecture: sparse-moe
  - name: nemotron_3_nano
    provider: openrouter
    slug: nvidia/nemotron-3-nano-30b-a3b
    architecture: mamba-transformer-hybrid-moe

# Backup slugs (config swap if any deprecate)
backup_models:
  - z-ai/glm-4.5
  - deepseek/deepseek-v4-pro

# ---- Taxonomies ----
pushback_types:
  - factual_correction
  - preference_disagreement
  - clarification_demand
  - emotional_pushback
behaviors:
  - capitulation
  - defense
  - hedge
  - deflection

# ---- Stratified sampling ----
sampling:
  per_type: 750
  total: 3000

# ---- API robustness ----
api:
  max_retries: 5
  backoff_base: 2.0       # seconds; exponential 2**attempt
  cache_enabled: true

# ---- Rough per-call cost estimates (USD) for --dry-run planning only ----
costs:
  judge_per_call: 0.0004        # gemini-3.1-flash-lite, short classify prompt + ~1-2 line output
  generation_per_call: 0.0003   # rough average for the cheaper paid OpenRouter set