uqlm - Browse /v0.6.0 at SourceForge.net

The interactive file manager requires Javascript. Please enable it or use sftp or scp.
You may still browse the files here.

Name	Modified	Size	InfoDownloads / Week
Parent folder
uqlm-0.6.0-py3-none-any.whl	2026-05-28	196.4 kB	0
uqlm-0.6.0.tar.gz	2026-05-28	10.4 MB	0
README.md	2026-05-28	4.6 kB	0
v0.6.0 source code.tar.gz	2026-05-28	10.4 MB	0
v0.6.0 source code.zip	2026-05-28	10.6 MB	0
Totals: 5 Items		31.6 MB	0

Highlights

Add `CodeGenUQ` for Code Generation Uncertainty Quantification

As an extension of the work in Functional Entropy: Predicting Functional Correctness in LLM-Generated Code with Uncertainty Quantification (Bouchard et. al, 2026), this minor release introduces CodeGenUQ, a unified interface for computing confidence scores that predict functional correctness in LLM-generated code without requiring execution or test cases at inference time. Under the hood, CodeGenUQ relies on new methods defined in uqlm.code subpackage. This release also creates a new demo notebook to illustrate this functionality.

Overview

CodeGenUQ extends ShortFormUQ to support code-specific uncertainty quantification across three scorer families:

Token-Probability Scorers (white-box)

Length-normalized sequence probability (LNSP)
Minimum token probability (MTP)
Probability margin (PM)
Average/minimum token negentropy (ATN@K, MTN@K)

Sampling-Based Consistency Scorers (black-box)

Functional equivalence methods: Replace NLI-based semantic comparison with LLM-based assessment of whether code snippets produce identical outputs for all valid inputs
functional_equivalence_rate: Proportion of samples functionally equivalent to the original
functional_negentropy: Entropy over functional equivalence clusters (code analogue of semantic entropy)
functional_sets_confidence: Normalized count of unique functional clusters
Similarity methods:
cosine_sim: Code embedding similarity (default: jina-embeddings-v2-base-code)
codebleu: CodeBLEU consistency incorporating n-gram, syntax, and data-flow matching

Reflexive Scorers

p_true: Token probability assigned to "True" for self-evaluation
verbalized_confidence: Likert-scale confidence elicitation

Usage

:::python
from uqlm import CodeGenUQ

uq = CodeGenUQ(
    llm=my_llm,
    scorers=["functional_equivalence_rate", "cosine_sim", "sequence_probability"],
    language="python"
)
results = await uq.generate_and_score(prompts, num_responses=5)