PromptForge-Quality

Multi-dimension prompt quality scorer. Given an LLM prompt, returns an overall quality score plus per-dimension scores, inferred issues, and missing information.

Part of PromptForge — local-first prompt scoring and optimization.

Model Details

Model Description

PromptForge-Quality is a fine-tuned answerdotai/ModernBERT-base encoder with regression heads that predict prompt quality on a 0–100 scale across seven dimensions.

• Developed by: PromptForge contributors
• Model type: Encoder + multi-output regression (promptforge_quality)
• Language(s): English
• License: MIT
• Finetuned from: answerdotai/ModernBERT-base (~150M parameters)

Dimensions scored

Dimension	What it measures
clarity	How clear and unambiguous the prompt is
specificity	Level of concrete detail
context	Background / situation provided
goal_definition	How well the objective is defined
constraints	Limits, requirements, must/must-not rules
completeness	Whether enough information is present
actionability	How easy it is for an LLM to act on
quality_score	Aggregate overall score

The model also surfaces issues (e.g. too_vague, missing_context) and missing_information hints.

Model Sources

• Repository: https://github.com/arjun988/promptModel
• Companion model: PromptForge-Optimizer (Qwen2.5-1.5B LoRA prompt rewriter)
• Demo: Gradio app in the PromptForge repo (demo/app.py)

Uses

Direct Use

• Score prompts before sending them to an LLM
• Diagnose weak prompts (what’s missing / unclear)
• Measure before/after quality when rewriting prompts
• Local / offline tooling via PromptForge CLI and Python API

Downstream Use

• Prompt linters in IDEs and agent frameworks
• Dataset filtering / ranking for synthetic prompt corpora
• Paired with PromptForge-Optimizer for score → optimize → re-score pipelines

Out-of-Scope Use

• Not a content moderator or safety classifier
• Not a judge of factual correctness of LLM answers
• Scores are calibrated on synthetic prompt quality labels — treat them as a useful proxy, not ground truth for every domain

Bias, Risks, and Limitations

• Trained largely on synthetic prompts with heuristic quality labels
• May reward length / structure more than true semantic quality
• English-centric; behavior on other languages is unverified
• Very domain-specific jargon may score inconsistently

Recommendations

• Use scores comparatively (before vs after) rather than as absolute grades
• Combine with human review for high-stakes prompt design
• For custom domains, retrain with your own labeled prompts

How to Get Started with the Model

Install & use with tuneprompt (recommended)

pip install tuneprompt

python -m promptforge download \
  --quality-repo ArjunShukla/PromptForge-Quality \
  --optimizer-repo ArjunShukla/PromptForge-Optimizer

python -m promptforge analyze "Build me a website"
# or: tuneprompt analyze "Build me a website"

from promptforge import PromptForge

pf = PromptForge(quality_model_path="ArjunShukla/PromptForge-Quality")
print(pf.analyze("Make an app."))
# → quality_score, dimensions, issues, missing_information

Package: tuneprompt on PyPI · Import: promptforge · CLI: tuneprompt / promptforge · Code: https://github.com/arjun988/promptModel

Full pipeline (score + optimize)

from promptforge import PromptForge

pf = PromptForge(
    quality_model_path="ArjunShukla/PromptForge-Quality",
    optimizer_model_path="ArjunShukla/PromptForge-Optimizer",
)
result = pf.run("Make an app about social media like facebook and stuff")
print(result["before"]["quality_score"], "→", result["after"]["quality_score"])
print(result["optimized_prompt"])

Training Details

Training Data

• ~25,000 synthetic prompts across coding, writing, research, data, and creative tasks
• Quality levels from vague one-liners to fully specified prompts
• Labels: overall quality_score + seven dimension scores (0–100)

Training Procedure

Setting	Value
Base model	answerdotai/ModernBERT-base
Task	Multi-dimension regression
Epochs	3
Max length	512
Precision	fp16
Hardware	NVIDIA RTX 5060 Laptop (8 GB)
Wall time	~33 minutes

Config: configs/quality_scorer.yaml

Evaluation

Held-out results (local training run):

Split	MAE	Pearson
Validation	2.73	0.993
Test (overall)	0.96	0.999

Test Spearman (overall): 0.959

Summary

Strong correlation with synthetic quality labels on held-out data. Real-world prompts should still be sanity-checked — the scorer is best used for ranking and diagnosing structure gaps.

Environmental Impact

• Hardware Type: NVIDIA RTX 5060 Laptop (8 GB)
• Hours used: ~0.5 h for this checkpoint
• Cloud Provider: N/A (local)
• Carbon Emitted: Not measured

Technical Specifications

Model Architecture and Objective

ModernBERT encoder with dual / multi regression heads predicting continuous quality scores (0–100).

Compute Infrastructure

• Hardware: RTX 5060 Laptop GPU, 8 GB VRAM
• Software: PyTorch (CUDA), Transformers, PromptForge training scripts

Artifact size

• On-disk checkpoint: ~574 MB

Citation

@software{promptforge_quality,
  title = {PromptForge-Quality},
  author = {PromptForge Contributors},
  year = {2026},
  url = {https://huggingface.co/ArjunShukla/PromptForge-Quality}
}

Model Card Contact

Open an issue on the PromptForge GitHub repository.