Add model card

Add comprehensive model card with training details, usage instructions, and links to blog posts

README.md

@@ -0,0 +1,92 @@
+---
+language:
+  - en
+license: apache-2.0
+library_name: transformers
+tags:
+  - ruby
+  - rails
+  - code-generation
+  - gguf
+  - fine-tuned
+  - lora
+  - unsloth
+pipeline_tag: text-generation
+base_model: Qwen/Qwen3-Coder-30B-A3B-Instruct
+model-index:
+  - name: qwen3-coder-30b-rails
+    results: []
+---
+
+# qwen3-coder-30b-rails
+
+A 31B parameter Mixture-of-Experts model fine-tuned for **Ruby on Rails code generation**. Trained on 111,000 samples extracted from 45 Rails repositories — 35 private client projects and 10 open-source codebases.
+
+Built by [Bytecode](https://bytecode.hr).
+
+## Model Details
+
+| Property | Value |
+|---|---|
+| Base model | [Qwen3-Coder-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct) |
+| Architecture | Qwen3 MoE (31B total, 3B active) |
+| Training method | QLoRA (rank 16) via [Unsloth](https://github.com/unslothai/unsloth) |
+| Training data | 111K samples from 45 Rails repos |
+| Training cost | ~$32 (A100 80GB, ~26 hours) |
+| Quantization | GGUF Q4_K_M (18.6 GB), Q5_K_M (21.7 GB) |
+
+## What it does
+
+This model writes idiomatic Ruby on Rails code following specific conventions:
+
+- Custom authentication with Identity and MagicLink models (not Devise)
+- Namespaced concerns instead of service objects
+- Solid Queue instead of Sidekiq
+- State-as-records instead of boolean flags
+- DaisyUI drawer layouts instead of ActiveAdmin
+
+It generates code that follows these patterns without prompt engineering — the conventions are baked into the weights.
+
+## Usage with Ollama
+
+```bash
+# Download and run
+ollama run bytecodehr/qwen3-coder-30b-rails
+
+# Example prompt
+ollama run bytecodehr/qwen3-coder-30b-rails "Write a Rails controller for managing user subscriptions with state transitions"
+```
+
+### Memory requirements
+
+| Format | GGUF Size | Min RAM | Recommended |
+|---|---|---|---|
+| Q5_K_M | 21.7 GB | 24 GB | 32 GB |
+| Q4_K_M | 18.6 GB | 20 GB | 24 GB |
+
+Rule of thumb: GGUF file size + 2–4 GB for KV cache and overhead.
+
+## Training
+
+Trained with LoRA (rank 16, alpha 16) on attention projection layers (`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`). Only 0.78% of parameters were trained.
+
+The dataset pipeline:
+1. Extracted code from 45 Rails repos (35 private + 10 open-source)
+2. 15-step cleaning and deduplication pipeline
+3. 111K final training samples
+4. Includes 29 contrastive pairs (wrong way vs right way)
+5. Source diversity cap at 20% per repository
+
+Full details in our blog posts:
+- [Part 1: Dataset Engineering](https://bytecode.hr/posts/training-rails-llms-part-1-dataset-engineering)
+- [Part 2: Training, Quantization, and Deployment](https://bytecode.hr/posts/training-rails-llms-part-2-training-quantization-deployment)
+
+## Why Ruby for LLMs?
+
+Ruby uses 42–45% fewer tokens than TypeScript across every major LLM tokenizer. That means more code fits in the context window, generations are faster, and costs are lower. Read our analysis: [Why Ruby Is the Better Language for LLM-Powered Development](https://bytecode.hr/posts/why-ruby-is-the-better-language-for-llm-powered-development).
+
+## Other models
+
+- [bytecodehr/qwen3-8b-rails](https://huggingface.co/bytecodehr/qwen3-8b-rails) — 8B dense model, runs on laptops (5 GB)
+- [bytecodehr/qwen2.5-coder-7b-rails](https://huggingface.co/bytecodehr/qwen2.5-coder-7b-rails) — 7B LoRA adapter
+- [bytecodehr/qwen2.5-coder-3b-rails](https://huggingface.co/bytecodehr/qwen2.5-coder-3b-rails) — 3B LoRA adapter