prism-creative-1-5-mini

<div align="center"> <img src="logo.png" alt="Prism logo" width="220"/>

Prism Creative 1.5 Mini

A compact local writing specialist -- creation, editing, and revision, distilled from five teacher models into Qwen3-4B

</div>

---

Overview

Prism Creative 1.5 Mini expands on 1 Mini's story-generation focus into a broader writing specialist: creative prose, poetry, editing, rewriting, revision, tone control, style transformation, and tightly constrained writing. Fine-tuned via LoRA on Qwen3-4B-Instruct-2507, the same base as 1 Mini, on a dataset distilled from five different teacher models rather than one, so the model learns the shared underlying skill instead of imitating a single teacher's voice.

Part of the Prism family of creative and roleplay models.

Training

• Base model: mlx-community/Qwen3-4B-Instruct-2507-4bit
• Teachers: a mix of nvidia/nemotron-3-nano-30b-a3b, deepseek-ai/deepseek-v4-flash-0731, nvidia/nemotron-3.5-lightning-30b-a3b, nvidia/nemotron-3-ultra-550b-a55b, and z-ai/glm-5.2 (all via NVIDIA NIM's free tier), blended for stylistic diversity
• Dataset: ~5,000 examples across 11 categories -- constrained prose, free-verse and formal poetry, grammar/clarity editing, tone-shift and tightening rewrites, restructuring, preserve/forbid-phrase constraints, style transfer, multi-constraint tasks, and correction-vs-complete-rewrite pairs
• Method: LoRA fine-tuning (rank 8, 16 layers), 12,000 iterations
• Best validation loss: 0.534 (reached at iteration 2,200 -- this checkpoint was used for fusing/publishing rather than the final iteration, since val loss plateaued and oscillated without further improvement for the rest of the run)

Evaluation

A held-out base-vs-tuned comparison (22 fresh scenarios spanning all 11 categories) shows a clear, measurable gain in the target skills:

The base model was already clean of reasoning leaks (this is Qwen3's Instruct variant), so the real gains are in conciseness and constraint discipline: on editing/tightening tasks the base model tends to ramble even when asked to cut a passage down, while the tuned model reliably produces a genuinely tightened result. On constraint tasks (e.g. "the rewrite must include this exact phrase"), the tuned model follows the instruction consistently; the base model is inconsistent about it. Repetition dropped substantially too. Poetry quality is comparable between the two -- both produce coherent, well-formed verse -- with the tuned model trending slightly more controlled and image-focused versus the base's more ornate tendency.

Formats available

This repo includes both:

Usage -- MLX

from mlx_lm import load, generate

model, tokenizer = load("VertexAGI/prism-creative-1-5-mini")

prompt = "Rewrite the following passage to tighten it by 40% while keeping every essential detail:\n\n[your passage here]"
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

response = generate(model, tokenizer, prompt=text, max_tokens=500)
print(response)

Usage -- GGUF (llama.cpp)

Like other Qwen3-based Prism models, the chat template only suppresses the model's reasoning trace when the runtime explicitly signals non-thinking mode. mlx-lm does this automatically; with llama.cpp pass -rea off (or --reasoning off), otherwise you'll see a <think>...</think> block before the actual output even though the underlying behavior is identical:

llama-cli -hf VertexAGI/prism-creative-1-5-mini -m prism_creative_1_5_mini_Q4_K_M.gguf -st -rea off -sys "You are an elite writing specialist -- equally skilled at creative prose, poetry, editing, rewriting, and tightly constrained writing tasks. Follow every requirement in the prompt precisely: length, tone, point of view, form, preserved/forbidden phrases, and any editing instructions. When editing or rewriting existing text, respect the specific instruction -- a light correction pass should change little, a complete rewrite should change freely, but either way preserve meaning unless told otherwise." -p "Write a haiku about dawn."

System prompt

For best results, use the system prompt the model was trained with:

You are an elite writing specialist -- equally skilled at creative prose, poetry, editing, rewriting, and tightly constrained writing tasks. Follow every requirement in the prompt precisely: length, tone, point of view, form, preserved/forbidden phrases, and any editing instructions. When editing or rewriting existing text, respect the specific instruction -- a light correction pass should change little, a complete rewrite should change freely, but either way preserve meaning unless told otherwise.

Limitations

This is a 4B-parameter model fine-tuned via LoRA on a moderately sized (~5,000-example) dataset -- capable but not infallible. Like any distilled model it inherits some of its teachers' stylistic tendencies, and very unusual poetic forms or highly technical editing tasks may still trip it up occasionally. Treat outputs as a strong starting point, not a guaranteed-perfect final one.

License

Apache 2.0, inherited from the Qwen3 base model.