Add model card

README.md

@@ -1,9 +1,96 @@
 ---
-library_name: mlx
 license: apache-2.0
-license_link: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507/blob/main/LICENSE
-pipeline_tag: text-generation
+language:
+- en
+base_model: mlx-community/Qwen3-4B-Instruct-2507-4bit
 tags:
 - mlx
-base_model: mlx-community/Qwen3-4B-Instruct-2507-4bit
+- gguf
+- creative-writing
+- storytelling
+- lora
+- distillation
+- conversational
+- text-generation
+pipeline_tag: text-generation
+---
+
+<div align="center">
+  <img src="logo.png" alt="Prism Creative logo" width="220"/>
+
+  # Prism Creative 1 Mini
+
+  **A frontier creative-writing model, distilled from GLM 5.2 into Qwen3-4B**
+</div>
+
 ---
+
+## Overview
+
+Prism Creative 1 Mini is a compact creative-writing and storytelling model, fine-tuned via LoRA on **Qwen3-4B-Instruct** using a distilled dataset generated by **GLM 5.2**. It's built to write across genres, forms, and tones — short stories, character sketches, poetry, dialogue scenes, worldbuilding passages, flash fiction, and more — while staying immersive and avoiding the flat, generic "AI assistant" register.
+
+Part of the **Prism** family of creative and roleplay models.
+
+## Training
+
+- **Base model:** `mlx-community/Qwen3-4B-Instruct-2507-4bit`
+- **Teacher model:** GLM 5.2 (via NVIDIA NIM)
+- **Dataset:** 2,500 examples across 10 categories — short stories, character sketches, opening lines, dialogue scenes, worldbuilding, poetry, genre blends, POV exercises, flash fiction, and continuations
+- **Method:** LoRA fine-tuning (rank 8, 16 layers), 6,500 iterations
+- **Best validation loss:** 1.064
+
+### Data quality
+
+The training data went through two hardening passes after an initial run surfaced two failure modes: a tendency to loop/repeat on very short fixed-length forms (like 100-word drabbles), and an over-reliance on a handful of "safe" cliché details (e.g. a clock reading a specific time) across otherwise-unrelated prompts. Both the generation prompt and an automated filter (rejecting outputs with excessive 4-gram repetition or flagged cliché phrases) were tightened before the final dataset was generated, and the existing data was re-filtered to match.
+
+## Evaluation
+
+A held-out sanity check across flash fiction, worldbuilding, poetry, and dialogue produced coherent, well-formed, in-genre output with no leaked planning text and no repetition loops. Example (flash fiction, mystery noir, 100-word constraint):
+
+> *Rain slicked the city like oil. I found her in the alley behind the pawn shop, fingers curled around a rusted key. "You knew," she whispered, eyes hollow...*
+
+## Formats available
+
+This repo includes both:
+
+| Format | File | Notes |
+|---|---|---|
+| **MLX** (4-bit) | `model.safetensors` + config | For Apple Silicon via `mlx-lm` |
+| **GGUF** (Q4_K_M) | `prism_creative_1_mini_Q4_K_M.gguf` | For `llama.cpp` and compatible runtimes (LM Studio, Ollama, etc.) |
+
+### Usage — MLX
+
+```python
+from mlx_lm import load, generate
+
+model, tokenizer = load("VertexAIco/prism-creative-1-mini")
+
+prompt = "Write a short story opening about a lighthouse keeper who receives a letter from someone who shouldn't exist."
+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=400)
+print(response)
+```
+
+### Usage — GGUF (llama.cpp)
+
+```bash
+llama-cli -hf VertexAIco/prism-creative-1-mini -m prism_creative_1_mini_Q4_K_M.gguf -p "Write a short story opening about a lighthouse keeper who receives a letter from someone who shouldn't exist."
+```
+
+## System prompt
+
+For best results, use the system prompt the model was trained with:
+
+```
+You are a skilled creative writer capable of working across genres, forms, and tones. Follow the given writing prompt precisely -- respect any length, format, and style constraints exactly. Write only the requested creative piece: no preamble, no meta-commentary about your approach, no "Here's a story about..." framing. Begin directly with the piece itself.
+```
+
+## Limitations
+
+This is a 4B-parameter model fine-tuned via LoRA on a moderately sized (2,500-example) dataset — it's capable but not infallible. It can still occasionally repeat itself on very constrained short forms, and like any distilled model it inherits some of its teacher's stylistic tendencies. Treat outputs as a strong first draft, not a final one.
+
+## License
+
+Apache 2.0, inherited from the Qwen3 base model.