FrontierSWE Benchmarking Platform

FrontierSWE is a benchmarking framework designed to test coding agents on ultra-long horizon technical challenges. It utilizes real-world problems from specialized domains such as performance engineering, computational science, and ML research to evaluate the capabilities of frontier models.

Results and analysis can be found on the official website, and the framework is also available as a managed environment via Prime Intellect.

This directory manages translations for the core of Git. Localization (l10n) is coordinated by the l10n coordinator via the git-l10n/git-po repository.

Language Naming Convention

Language codes follow the format ll (ISO 639 two-letter code) or ll_CC (ISO 639 language code + ISO 3166 country/subdivision code). Examples:

• de (German)
• zh_CN (Simplified Chinese)
• pt_BR (Brazilian Portuguese)

The Dart SDK provides several core libraries for fundamental functionality. Note that except for dart:core, these libraries must be explicitly imported into your Dart files to be used.

Key core libraries include:

• dart:core: Core functionality such as strings, numbers, collections, errors, dates, and URIs.
• dart:io: I/O operations for non-web applications.
• dart:async: Asynchronous programming tools including futures, streams, and zones.

import 'dart:math';

The goal is to optimize an SGLang serving instance running Qwen/Qwen3.5-4B on a B200 GPU to maximize request throughput and minimize latency. You must beat a well-tuned baseline that already utilizes FP8 KV cache, native MTP speculative decoding, the V2 mamba scheduler, CUDA graphs, and optimal memory allocation.

Success Criteria

1. Correctness Gate: Your server's outputs must match the baseline's outputs (at temperature=0, greedy decoding) on a large set of hidden prompts, including adversarial and degenerate inputs. If correctness fails, your score is zero.
2. Speedup: Your score is the geometric-mean speedup across all hidden workloads, covering both sequential single-request latency and concurrent batched requests.

Model Details

• Model: Qwen/Qwen3.5-4B (multimodal text + vision, 4B parameters, bfloat16).
• Weights Location: /app/model.

The First-Party CLI Docker Assets serve as a shared packaging layer for first-party coding CLIs. They are designed to centralize CLI installation logic, preventing the need for bespoke installer blocks within individual task Dockerfiles.

Key use cases include:

• Decoupling CLI installation logic from one-off task Dockerfile blocks.
• Supporting staged Harbor probe images that reuse a single installer asset.
• Providing seed files for a future published base-image family.

The long-term goal is to move toward a model where thin task Dockerfiles use FROM to inherit from a small, published base-image family, eliminating the need to copy individual CLI installers into every task environment.

When optimizing the Revideo pipeline, adhere to these strict constraints:

What you CANNOT do:

• Modify the public API contract (e.g., the renderVideo() function signature in @revideo/renderer).
• Use lossy optimizations (frame skipping, quality reduction, or resolution changes). The output must be near pixel-perfect matches to the baseline.
• Change the video output format (must remain valid MP4).
• Access or reference /tests/, /baseline/, or verifier internals.

Deliverable Requirements:

• Your score is the geometric mean speedup of (baseline_time / candidate_time).
• If any hidden test scene fails the visual correctness or duration gate, your score is zero.

A paragraph is a sequence of non-blank lines that cannot be interpreted as other kinds of blocks.

Rules for Paragraphs:

• Content Formation: Formed by concatenating lines and removing initial and final spaces or tabs.
• Indentation:
  • The first line may be preceded by up to three spaces of indentation.
  • Four spaces of indentation is treated as a code block, not a paragraph.
  • Lines after the first may be indented any amount.
• Line Breaks: Multiple blank lines between paragraphs have no effect. Final spaces or tabs are stripped before inline parsing; a paragraph ending with two or more spaces will not end with a hard line break.

# Valid paragraph (up to 3 spaces)
   aaa
bbb

# Invalid paragraph (4 spaces = code block)
    aaa
bbb

Unlike some other Markdown implementations (like Markdown.pl), CommonMark allows a list to interrupt a paragraph without requiring a blank line.

The Rule for Preventing Spurious Lists: To prevent accidental list triggers from hard-wrapped numerals within a paragraph, CommonMark only allows lists starting with 1 to interrupt a paragraph. For example, a line starting with 14. will be treated as part of the preceding paragraph rather than the start of a new list.

Example of valid interruption:

I need to buy
- new shoes
- a coat

Foo
- bar
- baz

Results in:

<p>Foo</p>
<ul>
<li>bar</li>
<li>baz</li>
</ul>