Benchmark Almanac

Delulu

Qwen2.5-Coder-32B-Instruct 84.5% leads since May 2026
Not in model cards

Models fill missing code in repository contexts across seven programming languages. Per-sample containers check compilation or type checking, while separate metrics compare completions with valid and hallucinated references.

The paper reports 1,951 samples, while the repository lists 1,947 samples and 947 distinct contexts. A separate classification track checks hallucination detection. Compiler success does not establish functional correctness.

Made by
Microsoft
Released
May 2026
Kind
Function-level problems
Level
University
Scored
Checked automatically
Answers
Code
Tasks
All public
Measures
Capability
Language
English
In and out
Text, Code → Code
Interaction
Single turn
Items from
Real work and use
Refresh
One fixed set
Run by
Anyone
Strongest model tested
StarCoder 2 15B #224 on the capability index
Last model added
November 2024
Scores refreshed
November 20th, 2026

In the authors’ words

Large Language Models for code generation frequently produce hallucinations in Fill-in-the-Middle (FIM) tasks – plausible but incorrect completions such as invented API methods, invalid parameters, undefined variables, or non-existent imports. These failures pass superficial review yet introduce runtime errors. We introduce Delulu, a verified multi-lingual benchmark of 1,951 FIM samples across 7 languages and 4 hallucination types. Samples are curated through an adversarial pipeline: a frontier LLM generates plausible hallucinations, four diverse judge models evaluate them, embedding-based clustering mines progressively harder examples, self-contained Docker containers verify that golden completions compile while hallucinated variants produce the expected runtime error, and a final human-expert review removes any remaining biased or trivially decidable samples. We evaluate 11 open-weight FIM models from five families spanning 0.5B-32B parameters: a six-point Qwen2.5-Coder scaling slate, plus a cross-family slate (CodeLlama, DeepSeek-Coder-V2, StarCoder2). The strongest model reaches only 84.5% pass@1, no family exceeds 0.77 Edit Similarity, and every family produces hallucination-aligned completions on a non-trivial share of samples, confirming that the difficulty exposed by Delulu is task-intrinsic rather than family-specific. We release the benchmark, containers, and evaluation framework at https://github.com/microsoft/delulu.

Abstract of Delulu: A Verified Multi-Lingual Benchmark for Code Hallucination Detection in Fill-in-the-Middle Tasks (arXiv:2605.07024, CC0)
11
Entries in its paper, 5 models we track

Qwen2.5-Coder-32B-Instruct leads with 84.5% among models released by November 2024 Qwen2.5-Coder-32B-Instruct leads with 84.5% among models released by November 2024

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.

The newest model with a score here came out in November 2024, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 Qwen2.5-Coder-32B-Instruct 84.5%
  2. 2 DSCoder-V2-Lite-Instruct 82.7%
  3. 3 Qwen2.5-Coder-14B-Instruct 81.5%
  4. 4 Qwen2.5-Coder-3B-Instruct 76.7%
  5. 5 CodeLlama-13B-hf 71.9%
  6. 6 CodeLlama-7B-hf 71.1%
  7. 7 Qwen2.5-Coder-7B-Instruct 71.1%
  8. 8 Qwen2.5-Coder-1.5B-Instruct 64.7%
  9. 9 Qwen2.5-Coder-0.5B-Instruct 43.6%
  10. 10 StarCoder 2 15B 37.2%
  11. 11 StarCoder 2 7B 19.3%

Source: The Delulu paper

#EntryScoreSettingDateSource
1 Qwen2.5-Coder-32B-Instruct 84.5% Nov 2024 Official leaderboard
2 DSCoder-V2-Lite-Instruct 82.7% Official leaderboard
3 Qwen2.5-Coder-14B-Instruct 81.5% Nov 2024 Official leaderboard
4 Qwen2.5-Coder-3B-Instruct 76.7% Nov 2024 Official leaderboard
5 CodeLlama-13B-hf 13B 71.9% Official leaderboard
6 CodeLlama-7B-hf 7B 71.1% Official leaderboard
7 Qwen2.5-Coder-7B-Instruct 7B 71.1% Official leaderboard
8 Qwen2.5-Coder-1.5B-Instruct 1.5B 64.7% Official leaderboard
9 Qwen2.5-Coder-0.5B-Instruct 0.5B 43.6% Official leaderboard
10 StarCoder 2 15B 37.2% Feb 2024 Official leaderboard
11 StarCoder 2 7B 19.3% Feb 2024 Official leaderboard
#ModelScoreSettingDateSource
1 Qwen2.5-Coder-32B-Instruct 84.5% Nov 2024 Official leaderboard
2 Qwen2.5-Coder-14B-Instruct 81.5% Nov 2024 Official leaderboard
3 Qwen2.5-Coder-3B-Instruct 76.7% Nov 2024 Official leaderboard
4 StarCoder 2 15B 37.2% Feb 2024 Official leaderboard
5 StarCoder 2 7B 19.3% Feb 2024 Official leaderboard

Rankings by part

We read 4 parts from the benchmark’s paper, each ranking models on one part of the test, such as Import, Method and Parameter. Qwen2.5-Coder-32B-Instruct leads 3 of them, and Qwen2.5-Coder-14B-Instruct leads the other one. Arrows show how far a model moved from its overall rank among the models we track.

Import

3 models we track have a score here.

  1. 1 Qwen2.5-Coder-32B-Instruct 77.1%
  2. 2 Qwen2.5-Coder-14B-Instruct 73.6%
  3. 3 Qwen2.5-Coder-3B-Instruct 65.3%

Number one over time

Lab claims count from the day the card came out, independent runs from the model’s release.

Qwen
Benchmark paper Not tested since May 2026

Source: The Delulu paper

Scores over time

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2024 2025 Near the ceiling (85%) Qwen2.5-Coder-32B-Instruct StarCoder 2 15B 0% 20% 40% 60% 80% 100% 2024 2025 Near the ceiling (85%) Qwen2.5-Coder-32B-Instruct StarCoder 2 15B

Source: Official Delulu leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Coding

All 523 in Coding
  • SWE-bench Verified In use 500 human-validated GitHub issues from 12 Python repos. Agent writes a patch that must pass hidden tests. Claude Opus 4.7 83.5%
  • LiveCodeBench v6 In use Competitive programming problems collected after model cutoffs from LeetCode, AtCoder, Codeforces. Scored pass@1 on hidden tests. DeepSeek-V4-Pro 93.5%
  • SWE-bench Multilingual In use 300 GitHub-issue tasks across nine programming languages in SWE-bench format. Patch graded by repository tests. Claude Fable 5.2 96.6%
  • DeepSWE In use 113 original long-horizon engineering tasks across 91 repos and five languages, written from scratch. Hand-written verifiers accept alternative solutions. GPT-6 Astra 74.1%
  • SciCode In use Scientist-curated research coding problems split into subproblems across natural sciences. Code checked against test cases. Seed2.1 Pro 59.8%
  • Codeforces rating In use Model solves recent Codeforces contest problems. An Elo-style rating is computed from simulated contest performance. DeepSeek V4.1 Flash 3,471

Data audit

9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 2 years ago, in November 2024, for Qwen2.5-Coder-32B-Instruct. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
  • Pass Independent runs Scores from the benchmark’s paper (11 entries).
  • Pass Models scored 5 models we track, plus 6 other entries on its leaderboard.
  • Warning Newest score The newest came 2 years ago, in November 2024, for Qwen2.5-Coder-32B-Instruct.
  • Note Frontier models None of the top 10 has a score. The strongest tested is StarCoder 2 15B (#224).
  • Pass Headroom The best is 84.5%, under 85%, so it still separates models.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 5 rows for models we track and 6 other entries at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check