Polish model card: lead with Cipher features, credit Qwen as foundation

README.md

@@ -11,89 +11,154 @@ tags:
 - gguf
 - ciphercode
 - vscode
+- developer-tools
 library_name: gguf
 ---
 
 # CipherModel-1.5B
 
-> **The model behind CipherCode™ — the AI coding assistant that writes code the way YOU would.**
-> Closed-beta v0.1, by **Lila AI LLC**.
+> **Your IDE's new best friend.**
+> The model behind [CipherCode](https://huggingface.co/guhantech) — the AI coding assistant that learns *your* style, remembers *your* projects, and writes code in *your* voice.
+>
+> By **Lila AI LLC** · Closed beta v0.1
 
-This repository hosts the GGUF Q4_K_M quantization served by the [CipherCode VS Code extension](https://github.com/lila-ai-llc/ciphercode-vscode) (closed beta). It is built on top of [Qwen/Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct) and is suitable for inline code completion, refactor / explain / fix / docstring tasks, and short conversational coding chat.
+---
 
-## What's in this repo
+## What CipherCode Delivers
 
-| File | Size | Format |
-|---|---|---|
-| `CipherModel-1.5B-Q4_K_M.gguf` | ~1.07 GB | GGUF Q4_K_M (llama.cpp) |
+CipherCode isn't another generic completion plugin. It's a complete coding companion that lives natively inside VS Code and adapts to *you*.
+
+### Cipher Persona — Your Style, Learned
+
+The first time you open a workspace, CipherCode silently scans your code and detects:
+
+- Naming conventions (camelCase / snake_case / PascalCase)
+- Function style (arrow vs named declarations)
+- Async style (async/await vs `.then`)
+- Comment placement and verbosity
+- Indent size, semicolon preference, type-annotation density
+- Your most-used libraries and imports
+
+From that moment forward, every suggestion is generated to feel like *you* wrote it. Nothing leaves your machine — Persona lives entirely in VS Code's `globalState`.
+
+### Project Memory — Continuity That Actually Helps
+
+CipherCode remembers your project across sessions:
+
+| What's tracked | Where |
+|---|---|
+| Project summary (auto-detected from `package.json` / README) | `.vscode/cipher-memory.json` |
+| Project type (`node` / `python` / `other`) | local |
+| Top 10 most-edited files | local |
+| Architectural decisions you've made | local |
+| Last 20 chat messages | local |
+| Recurring patterns in your code | local |
+
+This context is injected into every prompt, so when you come back tomorrow, the model already knows what you're building.
+
+### Smart Commands
+
+Right-click anywhere in your editor:
+
+- **Explain Code** — clear summary of what's happening, even without a selection
+- **Refactor Code** — clean up while preserving your style
+- **Fix Bug** — find and patch issues, style-matched
+- **Add Comments** — comment in your voice
+- **Document This File** — language-aware doc comments (TSDoc / JSDoc / Google Python / Javadoc / XMLDoc / Doxygen / godoc / rustdoc / PHPDoc / YARD)
+- **Generate README from Project** — full README from your code structure
+
+Plus an inline chat sidebar with persistent history, code-block copy buttons, "Insert at cursor" actions, and a stop button that actually stops.
+
+### Privacy by Architecture
 
-## What this is
+- Code stays on your machine — only the snippet you act on hits inference
+- Persona never leaves your laptop
+- Project memory lives in your workspace, not a Lila AI server
+- Self-hostable on your own GCP if you want full ownership
+- No telemetry, no accounts, no subscription
 
-- **A redistribution of `Qwen2.5-Coder-1.5B-Instruct` in GGUF Q4_K_M format**, branded as CipherModel-1.5B for use in the CipherCode extension's closed beta.
-- **No fine-tuning has been applied yet at v0.1.** The "Cipher Persona" style adaptation that ships with CipherCode operates entirely at the system-prompt level, injecting the developer's detected style into every request — model weights are unchanged from base Qwen.
-- A future v0.2+ release of this repo will contain a true LoRA fine-tune merged into the base.
+---
+
+## Powered By
+
+Built on **[Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct)** — Alibaba's state-of-the-art open code model — quantized to **Q4_K_M** for efficient CPU inference and packaged for deployment via `llama.cpp`.
+
+The intelligence in CipherCode comes from layering Persona detection, Project Memory, and carefully designed prompt templates on top of a strong base. The CipherCode VS Code extension orchestrates all of it; this repo hosts the weights it serves.
+
+A LoRA fine-tune is on the roadmap for v0.2 — trained on real-world IDE workflow patterns collected during the closed beta.
+
+## Specifications
+
+| | |
+|---|---|
+| **Architecture** | Qwen2.5-Coder transformer |
+| **Parameters** | 1.5 B |
+| **Context window** | 32 K (production runs at 4 K for efficiency) |
+| **Quantization** | Q4_K_M |
+| **File size** | 1.07 GB |
+| **License** | Apache 2.0 — free for commercial use |
+| **Strong languages** | Python, JavaScript, TypeScript, Java, Go, Rust, C/C++ |
 
-## Usage
+## Quick Start
 
-### Via the CipherCode VS Code extension (recommended)
+### Easy path — install the VS Code extension
+
+If Lila AI sent you the closed-beta `.vsix`:
 
 ```bash
-# Friends of Lila AI: install the .vsix sent to you privately
 code --install-extension ciphercode-0.1.0.vsix
 ```
 
-The extension talks to a private Cloud Run endpoint that serves this model via `llama-server`. End users of the extension never need to download this GGUF themselves.
+Open VS Code. Welcome walkthrough opens automatically. Start typing. No setup, no token, no GCP.
 
-### Direct with llama.cpp
+### Hands-on path — run the model locally
 
 ```bash
-# Download the GGUF
-huggingface-cli download guhantech/CipherModel-1.5B CipherModel-1.5B-Q4_K_M.gguf --local-dir .
+# Pull the GGUF
+hf download guhantech/CipherModel-1.5B \
+  CipherModel-1.5B-Q4_K_M.gguf --local-dir .
 
-# Run llama-server
+# Serve with llama-server
 llama-server \
   -m CipherModel-1.5B-Q4_K_M.gguf \
   --host 0.0.0.0 --port 8080 \
   --ctx-size 4096 -np 5
 
-# Hit it
+# Make a request
 curl -X POST http://localhost:8080/v1/chat/completions \
   -H "Content-Type: application/json" \
-  -d '{"model":"cipher-model","messages":[{"role":"user","content":"write a python fizzbuzz"}],"max_tokens":256}'
+  -d '{
+    "model": "cipher-model",
+    "messages": [{"role":"user","content":"write a python fizzbuzz"}],
+    "max_tokens": 256
+  }'
 ```
 
-### Direct with `llama-cpp-python`
+### Python (`llama-cpp-python`)
 
 ```python
 from llama_cpp import Llama
+
 llm = Llama(model_path="CipherModel-1.5B-Q4_K_M.gguf", n_ctx=4096)
 out = llm("def fizzbuzz(n):", max_tokens=256)
 print(out["choices"][0]["text"])
 ```
 
-## Specifications
-
-- **Architecture:** Qwen2.5-Coder (transformer)
-- **Parameters:** 1.5 B
-- **Context window:** 32 K (we run at 4 K in production for memory)
-- **Quantization:** Q4_K_M
-- **License:** Apache 2.0 (inherited from base model)
-- **Languages supported:** strong in Python, JavaScript, TypeScript, Java, Go, Rust, C/C++ — see Qwen2.5-Coder's eval table for details
+## Roadmap
 
-## Limitations
-
-- Quality is meaningfully lower than Qwen-Coder-7B / 32B. For complex multi-file reasoning or long-context tasks, prefer the larger sizes.
-- Q4_K_M trades ~1–2% quality for ~4× smaller size vs full fp16. Acceptable for autocomplete and single-file tasks.
-- This is a closed-beta artifact; no SLAs, no support guarantees.
-
-## Citation / credits
+| Version | Status | What's in it |
+|---|---|---|
+| **v0.1** | Live | Closed beta. Cipher Persona + Project Memory + 11 commands + chat sidebar. |
+| **v0.2** | Planned | LoRA fine-tune on collected IDE workflows. Better instruction-following. |
+| **v0.3** | Planned | Multi-file context awareness. Whole-project doc generation. |
+| **v1.0** | Planned | Public Marketplace launch. Optional hosted Pro tier for zero-setup. |
 
-Built on top of:
+## Citation
 
 ```bibtex
 @article{hui2024qwen2,
   title={Qwen2.5-Coder Technical Report},
-  author={Binyuan Hui and Jian Yang and Zeyu Cui and Jiaxi Yang and Dayiheng Liu and Lei Zhang and Tianyu Liu and Jiajun Zhang and Bowen Yu and Keming Lu and Kai Dang and Yang Fan and Yichang Zhang and An Yang and Rui Men and Fei Huang and Bo Zheng and Yibo Miao and Shanghaoran Quan and Yunlong Feng and Xingzhang Ren and Xuancheng Ren and Jingren Zhou and Junyang Lin},
+  author={Hui, Binyuan and Yang, Jian and Cui, Zeyu and others},
   journal={arXiv preprint arXiv:2409.12186},
   year={2024}
 }
@@ -101,10 +166,10 @@ Built on top of:
 
 ## Trademark
 
-CipherCode™ and Cipher Persona™ are trademarks of **Lila AI LLC**. All rights reserved.
+**CipherCode** and **Cipher Persona** are trademarks of **Lila AI LLC**. All rights reserved.
 
-The CipherModel weights themselves are released under Apache 2.0 (inherited from Qwen). The trademarks restrict only how you may name and brand derivative work — the underlying weights are free to use.
+The model weights are released under Apache 2.0 — free to use, modify, and redistribute. Trademarks restrict only how you may name and brand derivative work; the underlying weights remain unrestricted.
 
 ---
 
-© 2026 Lila AI LLC.
+<sub>© 2026 Lila AI LLC · Built for developers who don't want their AI to sound like Stack Overflow.</sub>