Skip to content

[codex] Add atomic edit format and DeepSeek V4 Pro polyglot result - #5282

Open
danielgonzagat wants to merge 61 commits into
Aider-AI:mainfrom
danielgonzagat:bench/deepseek-v4-pro-polyglot-result
Open

[codex] Add atomic edit format and DeepSeek V4 Pro polyglot result#5282
danielgonzagat wants to merge 61 commits into
Aider-AI:mainfrom
danielgonzagat:bench/deepseek-v4-pro-polyglot-result

Conversation

@danielgonzagat

Copy link
Copy Markdown

Summary

This PR adds the atomic whole-file edit format work and supporting benchmark/runtime changes used to run an official polyglot benchmark at scale. It also adds the resulting DeepSeek V4 Pro entry to aider/website/_data/polyglot_leaderboard.yml.

The submitted leaderboard row reports:

  • 212/225 passing on try 2, 94.2%
  • 82/225 passing on try 1, 36.4%
  • model: DeepSeek V4 Pro
  • command: aider --model deepseek/deepseek-v4-pro
  • edit format: atomic
  • benchmarked commit: c4d9f23
  • date: 2026-06-18

Benchmark Evidence

The full run was executed as a Modal sharded polyglot run, one shard per language, using the same model, edit format, commit hash, and polyglot benchmark ref.

Shard summaries:

  • C++: 24/26
  • Go: 36/39
  • Java: 44/47
  • JavaScript: 46/49
  • Python: 33/34
  • Rust: 29/30

Aggregate failures: cpp/bank-account, cpp/meetup, go/connect, go/counter, go/robot-simulator, java/forth, java/rational-numbers, java/sgf-parsing, javascript/complex-numbers, javascript/list-ops, javascript/rational-numbers, python/forth, rust/forth.

Validation

Local verification on the PR branch:

  • .venv/bin/python -m pytest tests/basic/test_coder.py tests/basic/test_wholefile.py tests/basic/test_linter.py benchmark/test_benchmark.py benchmark/test_modal_runner.py -q -> 121 passed, 1 skipped, 29 subtests passed
  • .venv/bin/python -m py_compile aider/coders/base_coder.py tests/basic/test_coder.py benchmark/benchmark.py benchmark/modal_runner.py
  • git diff --check
  • YAML parse with yaml.safe_load()
  • independent aggregate cross-check from the 225 r11 per-case JSONs matched the leaderboard row: pass_num_2=212, pass_rate_2=94.2

bundle exec jekyll build was attempted locally but the local Ruby environment does not have jekyll installed (bundler: command not found: jekyll).

@CLAassistant

CLAassistant commented Jun 19, 2026

Copy link
Copy Markdown

CLA assistant check
All committers have signed the CLA.

@danielgonzagat
danielgonzagat marked this pull request as ready for review June 19, 2026 00:43
…ntation)

Removes the test-error reflection augmentation that injected hidden test source
into the model's retry prompt. base_coder.py now matches upstream; the benchmark
passes the raw test output through. Scores are comparable to the official harness.
@danielgonzagat
danielgonzagat force-pushed the bench/deepseek-v4-pro-polyglot-result branch from 17b385d to f39118c Compare June 19, 2026 02:33
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants