openenv.yaml

name: long-context-summarization
version: "1.0.0"
description: >
  An RL environment for training models to compress long documents into
  compact summaries and then use those summaries to answer questions.
  Inspired by Cursor's self-summarization approach (cursor.com/blog/self-summarization).
author: "Hackathon Team"

tasks:
  - name: easy
    description: "Short Wikipedia passage (300–700 chars). Truncated to 70%. Factual QA."
    difficulty: easy
    dataset: "rajpurkar/squad"
    max_steps: 2
    reward_range: [0.0, 1.0]
    grader: token_f1

  - name: medium
    description: "Longer Wikipedia passage (900–2500 chars). Truncated to 65%. Factual QA."
    difficulty: medium
    dataset: "rajpurkar/squad"
    max_steps: 2
    reward_range: [0.0, 1.0]
    grader: token_f1

  - name: hard
    description: >
      Long scientific paper (QASPER, 2000–10000 chars). Truncated to 55%,
      split into two chunks. Chained summarization (3 steps): summarize chunk1,
      update summary with chunk2, then answer.
    difficulty: hard
    dataset: "allenai/qasper"
    max_steps: 3
    reward_range: [0.0, 1.0]
    grader: token_f1

endpoints:
  reset: "POST /reset"
  step: "POST /step"
  state: "GET /state"
  health: "GET /health"
  schema: "GET /schema"

action_space:
  type: text
  description: "Free-form text response (summary or answer)"

observation_space:
  messages:
    type: "array of {role, content} dicts"
    description: "Full conversation history in OpenAI chat format"
  step_type:
    type: string
    values: ["summarize", "update_summary", "answer", "done"]
  task_name:
    type: string
    values: ["easy", "medium", "hard"]
  context_length:
    type: integer
    description: "Total context length in characters"
  truncation_ratio:
    type: float
    description: "Fraction of context shown to the model (0.55–0.70)"
  category:
    type: string
    description: "Coarse content domain such as history, science, geography, or scientific_research"
  source_type:
    type: string
    description: "Source style such as encyclopedic_passage, long_form_reference, or scientific_paper"
  done:
    type: boolean
  reward:
    type: float
    range: [0.0, 1.0]
    description: "Token-level F1 score (null until final step)"

reward_shaping:
  primary: "Token-level F1 score between predicted answer and ground-truth answers"
  conciseness_bonus: "Up to +0.05 for summaries under 150 words when the answer is correct"
  optional_llm_grader: "When USE_LLM_GRADER=true, blends deterministic answer grading with an auxiliary LLM summary-quality score"
  configurability: "Reward thresholds, bonuses, and hybrid weights can be tuned via GRADER_* environment variables"