| ---
|
| base_model: Qwen/Qwen2.5-1.5B-Instruct
|
| library_name: peft
|
| license: mit
|
| language:
|
| - en
|
| pipeline_tag: text-generation
|
| tags:
|
| - promptforge
|
| - prompt-optimization
|
| - prompt-engineering
|
| - lora
|
| - peft
|
| - qwen2.5
|
| - text-generation
|
| - base_model:adapter:Qwen/Qwen2.5-1.5B-Instruct
|
| - transformers
|
| ---
|
|
|
| # PromptForge-Optimizer
|
|
|
| LoRA adapter that rewrites **weak / vague prompts** into **clear, specific, actionable LLM prompts** while preserving the original intent and topic.
|
|
|
| Part of [PromptForge](https://github.com/arjun988/promptModel) — local-first prompt quality scoring + optimization.
|
|
|
| ## Model Details
|
|
|
| ### Model Description
|
|
|
| PromptForge-Optimizer is a **PEFT/LoRA** fine-tune of [`Qwen/Qwen2.5-1.5B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct). Given a weak user prompt (plus optional quality analysis context), it generates an improved prompt with audience, constraints, structure, and output format — without changing the core topic.
|
|
|
| - **Developed by:** PromptForge contributors
|
| - **Model type:** Causal LM adapter (LoRA / PEFT)
|
| - **Language(s):** English
|
| - **License:** MIT
|
| - **Finetuned from:** [`Qwen/Qwen2.5-1.5B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct)
|
|
|
| ### Model Sources
|
|
|
| - **Repository:** https://github.com/arjun988/promptModel
|
| - **Companion model:** PromptForge-Quality (ModernBERT multi-dimension prompt scorer)
|
| - **Demo:** Gradio app in the PromptForge repo (`demo/app.py`)
|
|
|
| ## Uses
|
|
|
| ### Direct Use
|
|
|
| - Rewrite vague prompts into production-ready LLM instructions
|
| - Pair with **PromptForge-Quality** for score → optimize → re-score workflows
|
| - Local / offline prompt tooling (CLI, Python API, Gradio)
|
|
|
| Example weak → strong:
|
|
|
| | Weak | Optimized (intent preserved) |
|
| |------|------------------------------|
|
| | `Make an app about social media like facebook and stuff` | Social media / Facebook-like app prompt with profiles, feed, likes, constraints, output format |
|
|
|
| ### Downstream Use
|
|
|
| - Prompt engineering assistants
|
| - IDE / agent tooling that improves user instructions before calling an LLM
|
| - Synthetic data pipelines that need higher-quality prompts
|
|
|
| ### Out-of-Scope Use
|
|
|
| - Not a general chat assistant
|
| - Not a substitute for domain experts (legal, medical, safety-critical advice)
|
| - Not guaranteed to preserve intent on topics far outside the curated training set
|
| - Do not use to generate harmful, deceptive, or disallowed content
|
|
|
| ## Bias, Risks, and Limitations
|
|
|
| - Trained on **curated synthetic** weak→strong pairs; coverage is strongest on coding apps, writing, data, research, and planning prompts
|
| - May invent plausible audience / stack details (e.g. “product managers”, “Flask”) when the weak prompt is underspecified
|
| - Small base model (**1.5B**) — quality is good for local use, not frontier-LLM rewrite quality
|
| - Inference includes validation + fallback in the PromptForge package; raw adapter output alone may still drift
|
|
|
| ### Recommendations
|
|
|
| - Prefer the **PromptForge Python package / CLI** (chat template + stop tokens + validation) over raw `generate`
|
| - For new domains, add your own weak→strong pairs and retrain the LoRA
|
| - Always review optimized prompts before sending them to production LLMs
|
|
|
| ## How to Get Started with the Model
|
|
|
| ### Install & use with [`tuneprompt`](https://pypi.org/project/tuneprompt/) (recommended)
|
|
|
| ```bash
|
| pip install tuneprompt
|
|
|
| python -m promptforge download \
|
| --quality-repo ArjunShukla/PromptForge-Quality \
|
| --optimizer-repo ArjunShukla/PromptForge-Optimizer
|
|
|
| python -m promptforge run "Make an app about social media like facebook and stuff"
|
| # or: tuneprompt run "Make an app about social media like facebook and stuff"
|
| ```
|
|
|
| ```python
|
| from promptforge import PromptForge
|
|
|
| pf = PromptForge(
|
| quality_model_path="ArjunShukla/PromptForge-Quality",
|
| optimizer_model_path="ArjunShukla/PromptForge-Optimizer",
|
| )
|
| print(pf.run("Build me a website for a startup")["optimized_prompt"])
|
| ```
|
|
|
| > **Package:** [`tuneprompt`](https://pypi.org/project/tuneprompt/1.0.0/) on PyPI · **Import:** `promptforge` · **CLI:** `tuneprompt` / `promptforge` · **Code:** https://github.com/arjun988/promptModel
|
|
|
| ### Load the adapter directly (PEFT)
|
|
|
| ```python
|
| from peft import PeftModel
|
| from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
| base = "Qwen/Qwen2.5-1.5B-Instruct"
|
| adapter = "ArjunShukla/PromptForge-Optimizer"
|
|
|
| tokenizer = AutoTokenizer.from_pretrained(adapter, trust_remote_code=True)
|
| model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True, device_map="auto")
|
| model = PeftModel.from_pretrained(model, adapter)
|
| ```
|
|
|
| Use Qwen’s chat template (`tokenizer.apply_chat_template`) — do not hand-roll `<|system|>` tags.
|
|
|
| ## Training Details
|
|
|
| ### Training Data
|
|
|
| - **~800** curated high-quality **weak → strong** prompt pairs
|
| - **~140** unique topic-preserving seeds (coding, writing, data, research, general)
|
| - Intent rule: optimized prompt must keep the same topic as the weak prompt
|
| - Assistant-only loss masking (system/user tokens not trained)
|
|
|
| ### Training Procedure
|
|
|
| #### Training Hyperparameters
|
|
|
| | Setting | Value |
|
| |---------|-------|
|
| | Base model | `Qwen/Qwen2.5-1.5B-Instruct` |
|
| | Method | LoRA (PEFT) |
|
| | LoRA rank / alpha | 16 / 32 |
|
| | Target modules | q/k/v/o + MLP projections |
|
| | Max sequence length | 512 |
|
| | Epochs | 6 |
|
| | Effective batch size | 8 (batch 1 × grad accum 8) |
|
| | Learning rate | 1e-4 |
|
| | Precision | fp16 |
|
| | Gradient checkpointing | enabled |
|
| | Config | `configs/optimizer_fast_8gb.yaml` |
|
|
|
| #### Speeds, Sizes, Times
|
|
|
| - **Hardware:** NVIDIA GeForce RTX 5060 Laptop GPU (8 GB)
|
| - **Wall time:** ~87 minutes (6 epochs)
|
| - **Adapter size on disk:** ~82 MB
|
| - **Train loss:** ~0.47
|
| - **Validation loss:** ~0.121
|
|
|
| ## Evaluation
|
|
|
| ### Metrics
|
|
|
| | Signal | Result |
|
| |--------|--------|
|
| | Validation loss | **0.121** |
|
| | Example quality lift (scorer) | e.g. **41.5 → 94.0** on a social-media app prompt |
|
| | Intent preservation | Topic keywords retained (social / Facebook) |
|
| | Validation gate | Rejects empty / repetitive / low-intent outputs |
|
|
|
| Evaluation is primarily: held-out SFT loss + pipeline checks (score delta, instruction preservation, repetition detection). Not a public leaderboard benchmark.
|
|
|
| ### Summary
|
|
|
| The adapter reliably expands vague prompts into structured instructions on in-distribution topics. Off-distribution prompts may fall back to a safer template when used through PromptForge.
|
|
|
| ## Environmental Impact
|
|
|
| - **Hardware Type:** NVIDIA RTX 5060 Laptop (8 GB)
|
| - **Hours used:** ~1.5 h for this adapter run
|
| - **Cloud Provider:** N/A (local)
|
| - **Compute Region:** N/A
|
| - **Carbon Emitted:** Not measured
|
|
|
| ## Technical Specifications
|
|
|
| ### Model Architecture and Objective
|
|
|
| - Causal language model (Qwen2.5 Instruct) + LoRA
|
| - Objective: SFT to map weak prompt (+ analysis) → optimized prompt text only
|
|
|
| ### Compute Infrastructure
|
|
|
| #### Hardware
|
|
|
| - RTX 5060 Laptop GPU, 8 GB VRAM
|
|
|
| #### Software
|
|
|
| - PyTorch (CUDA)
|
| - Transformers
|
| - PEFT / LoRA
|
| - PromptForge training scripts
|
|
|
| ### Framework versions
|
|
|
| - PEFT 0.20.0
|
|
|
| ## Citation
|
|
|
| ```bibtex
|
| @software{promptforge_optimizer,
|
| title = {PromptForge-Optimizer},
|
| author = {PromptForge Contributors},
|
| year = {2026},
|
| url = {https://huggingface.co/ArjunShukla/PromptForge-Optimizer}
|
| }
|
| ```
|
|
|
| ## Model Card Contact
|
|
|
| Open an issue on the PromptForge GitHub repository.
|
|
|
