---

base_model: unsloth/Qwen3.5-2B

tags:

• gguf
• ollama
• llama.cpp
• qwen3.5
• prompt-engineering

pipeline_tag: text-generation

---

Qwen3.5-2B Prompt Compiler (GGUF)

Quantized GGUF build of a LoRA fine-tune of unsloth/Qwen3.5-2B, for use with Ollama or llama.cpp.

The adapter (safetensors, for transformers/peft) is at tshaik1990/qwen3.5-2b-prompt-compiler.

Files

| File | Quantization | Size |

|---|---|---|

| Qwen3.5-2B.Q4_K_M.gguf | Q4_K_M | ~1.31 GB |

| Qwen3.5-2B.F16-mmproj.gguf | F16 (vision projector) | ~668 MB |

SHA256 checksums are in SHA256SUMS.txt — verify with shasum -a 256 -c SHA256SUMS.txt.

What it does

Given a system prompt instructing it to act as a prompt compiler, it turns a short, underspecified request into a clear, structured, implementation-ready prompt — asking for missing context (budget, location, constraints, etc.) instead of inventing it, and staying in the "write a prompt" role instead of drifting into answering the request directly.

Use with Ollama

ollama run tshaik1990/qwen3.5-2b-prompt-compiler-GGUF

Or build locally from a downloaded GGUF with a Modelfile:

FROM ./Qwen3.5-2B.Q4_K_M.gguf
FROM ./Qwen3.5-2B.F16-mmproj.gguf

RENDERER qwen3.5
PARSER qwen3.5

SYSTEM """You are a prompt compiler. Convert the user's request into a clear, structured, implementation-ready prompt while preserving all stated requirements."""

PARAMETER temperature 0.2
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER num_ctx 4096

ollama create qwen3.5-2b-prompt-compiler -f Modelfile
ollama run qwen3.5-2b-prompt-compiler

Evaluation

Benchmarked against the unmodified base model under identical sampling settings (temperature 0.2, top_p 0.95, top_k 20, same system prompt) across 20 held-out prompts:

• Average output length: 2775 chars (base) vs 1002 chars (fine-tuned)
• Leaks meta-commentary ("Here is a prompt for...") instead of just emitting the prompt: 55% (base) vs 0% (fine-tuned)
• Opens with a role assignment ("You are an expert..."): 0% (base) vs 85% (fine-tuned)
• Shorter output on 18 of 20 prompts

Limitations

Trained for a single epoch on a small, self-authored dataset; the evaluation above is self-authored, not an independent benchmark, and human-preference review is still pending. Small (2B) model — may not generalize well outside prompt-compiling use cases.

License

Inherits the license of the base model, unsloth/Qwen3.5-2B.