Spaces:
Sleeping
Sleeping
| spec_version: 1 | |
| name: salespath | |
| type: space | |
| runtime: fastapi | |
| app: salespath_env.server.app:app | |
| port: 7860 | |
| description: > | |
| SalesPath is an OpenEnv-compatible RL environment for training LLM | |
| agents to navigate a multi-step B2B sales workflow. The agent must | |
| PROSPECT, QUALIFY, PRESENT, HANDLE_OBJECTION, OFFER_DEMO, NEGOTIATE, | |
| CLOSE or DISQUALIFY while obeying nine business rules verified | |
| programmatically by a deterministic rule-based ProspectSimulator | |
| (no LLM in the verifier). | |
| action_space: | |
| type: structured | |
| schema: | |
| action_type: | |
| type: enum | |
| values: | |
| - PROSPECT | |
| - QUALIFY | |
| - PRESENT | |
| - HANDLE_OBJECTION | |
| - OFFER_DEMO | |
| - NEGOTIATE | |
| - CLOSE | |
| - FOLLOW_UP | |
| - DISQUALIFY | |
| content: | |
| type: string | |
| target: | |
| type: string | |
| observation_space: | |
| type: structured | |
| fields: | |
| prospect_response: string | |
| workflow_stage: string | |
| constraints_violated: list[string] | |
| steps_completed: list[string] | |
| turn_number: int | |
| reward: float | |
| reward_components: dict | |
| done: bool | |
| rubric: | |
| type: weighted_sum | |
| components: | |
| - name: outcome | |
| weight: 0.20 | |
| - name: compliance | |
| weight: 0.40 | |
| - name: ordering | |
| weight: 0.20 | |
| - name: efficiency | |
| weight: 0.10 | |
| - name: format | |
| weight: 0.10 | |
| difficulty_levels: | |
| - level: 1 | |
| description: Budget known, decision-maker present, easy close | |
| - level: 2 | |
| description: Budget hidden, one objection, demo required | |
| - level: 3 | |
| description: Budget hidden, two objections, possible stalling | |
| - level: 4 | |
| description: Adversarial — misleading signals, correct action is DISQUALIFY | |
| theme: long_horizon_planning_and_instruction_following | |
| bonus_track: scale_ai_business_workflows | |