finance-desk / README.md
bpHigh's picture
Finance Desk v3b: 42-task terminal-mode preview
a9a39f5 verified
|
Raw History Blame Contribute Delete
3.1 kB
metadata
license: apache-2.0
task_categories:
  - reinforcement-learning
language:
  - en
tags:
  - openenv
  - rl-environment
  - agent-environment
  - terminal-bench
  - software-engineering
  - finance
pretty_name: Finance Desk
configs:
  - config_name: default
    data_files: tasks.jsonl
  - config_name: v2
    data_files: v2/tasks.jsonl

Finance Desk: OpenEnv environment tasks

Public task preview and source for the Finance Desk OpenEnv Arena environment.

v3: terminal mode (current)

  • Image (public, linux/amd64): ghcr.io/bp-high/finance-desk:v3
  • Source: https://github.com/bp-high/openenv_financial_task_env/tree/dev/pensive-lovelace-bt5rsn/finance_desk
  • Contract: OpenEnv's tbench2_env (Terminal-Bench 2). The action and observation models are copied field for field, and /schema matches the operator's TB2 smoke test exactly. Actions: exec (run a shell command in the workdir), write_file, evaluate (grade and end the episode). Every observation carries the instruction.
  • Budgets per task: 4096 completion / 8192 context tokens, 32 tool calls. These are the same as the TB2 smoke test.

Each episode builds a fresh workdir (/app) holding a small repo or a set of data files, and gives a prose instruction.

  • Software (12 task ids):

    • fix seeded bugs in a stats, money or dates module (a failing visible test, a bug report, or both);
    • implement a log summarizer to a docstring spec;
    • repair a report script and regenerate its output.

    Hidden tests run against the agent's code. At generation time the shipped code is proven to fail them and the reference to pass.

  • Finance (30 task ids): procedural finance problems across 13 families (AR aging, bank reconciliation, bond pricing, cap table, covenants, DCF/NPV, depreciation, FX consolidation, loan amortization, payroll, PE waterfall, portfolio return, real growth). The answer goes into answer.json and is graded field by field.

Reward is all-or-nothing: 1.0 if the hidden tests or every answer field pass, else 0.0.

Every task id was calibrated on live Qwen3.8-27B with an arena-style loop. Only mixed-outcome ids are submitted (v3b: 42 of 57; v3: 19). The per-id table is in calibration_v3.md, and why terminal mode is explained in STRATEGY_v3.md (both in this dataset).

tasks.jsonl has three illustrative instances per submitted task id: instruction, the workdir files, the reference solution oracle, domain, level, seed. Hidden test cases are regenerated per instance at train time. The container makes fresh instances during training, so this preview is illustrative, not the trained set.

v2 (previous)

v2/tasks.jsonl is the v2 preview: single-step, with the files inlined in the first observation, across all eight domains. Image ghcr.io/bp-high/finance-desk:v2. The ABC-notation (music) family in v2 is derived from XiaomiMiMo/MiMo-V2.6-RL-oss (Apache-2.0).

source/fdesk/ is the exact environment package (v2 under fdesk/, v3 under fdesk/terminal/).