Create README.md

README.md

@@ -0,0 +1,80 @@
+---
+model-name: Dirty-Calla (4B)
+license: other
+base_model: Daizee/Gemma3-Callous-Calla-4B
+library_name: transformers
+pipeline_tag: text-generation
+language:
+- en
+tags:
+- gemma
+- gemma-3
+- 4b
+- sft
+- lora
+- lora-merged
+- gguf
+- llama.cpp
+- text-generation
+- creative-writing
+datasets:
+- private
+---
+
+# Dirty-Calla (4B)
+
+**Dirty-Calla** is a light **style SFT** of `Daizee/Gemma3-Callous-Calla-4B`. It was trained on a dataset of fanfiction with user prompts (eg Write me a story with XXX character and XXX theme.) Initial data was synthetically increased from ~360 to 3000. The idea is to prompt Dirty Calla to write short fictional stories with tags and general, big picture idea.
+
+This repo contains:
+- **Merged Transformers weights** (`./`): LoRA merged into the base for easy inference with `transformers`.
+- **LoRA adapters** (`./adapters/`): for reproducibility or further fine-tuning.
+- **GGUF builds** (`./gguf/`): for fast, local inference via `llama.cpp` runtimes.
+
+> **Note:** You are responsible for the content you generate and how you distribute it. Follow all applicable laws and platform policies. Respect the **Gemma** license and Hugging Face terms.
+
+---
+
+## 🧱 Base & Provenance
+
+- **Base model:** `Daizee/Gemma3-Callous-Calla-4B` (merged from Gemma-3 4B IT derivatives via TIES; tokenizer from the Gemma-3 family).
+- **Architecture:** Gemma-3 4B (instruction-tuned).
+- **Training method:** Light **LoRA SFT** (single epoch by default) on curated, style-targeted dialogs; adapters merged into base after training.
+
+---
+
+## 📦 Files
+
+### Transformers (merged)
+- `config.json`, `model-00001/2-of-00002.safetensors`, `tokenizer.json`, `tokenizer.model` (SPM), etc.
+
+### Adapters (optional)
+- `adapters/checkpoint-*/adapter_model.safetensors`, `adapter_config.json`, etc.
+
+### GGUF (llama.cpp)
+- `gguf/dirty-calla-q8_0.gguf` — near-lossless quality.
+- `gguf/dirty-calla-q5_k_m.gguf` — **recommended** quality/speed sweet spot.
+- `gguf/dirty-calla-q4_0.gguf` — compact “plain q4”.
+- *(optional)* `gguf/dirty-calla-q4_k_m.gguf` — smaller with improved quality vs q4_0 in many cases.
+
+> If you don’t see a particular quant, it may not have been uploaded yet.
+
+---
+
+## 🚀 Quick Start (Transformers)
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+MODEL = "Daizee/Dirty-Calla"
+tok = AutoTokenizer.from_pretrained(MODEL, use_fast=True)
+model = AutoModelForCausalLM.from_pretrained(MODEL, device_map="auto")
+
+# Gemma-3 style chat template (example)
+dialog = [
+    {"role": "system", "content": "You are Dirty-Calla, a bold, stylish fiction writer. Be vivid and punchy."},
+    {"role": "user", "content": "Give me a one-paragraph teaser for a dramatic, slow-burn romance. Keep it PG-13."}
+]
+prompt = tok.apply_chat_template(dialog, tokenize=False, add_generation_prompt=True)
+inputs = tok(prompt, return_tensors="pt").to(model.device)
+out = model.generate(**inputs, max_new_tokens=220, temperature=0.9, top_p=0.9)
+print(tok.decode(out[0], skip_special_tokens=True))