Frank 9B
Frank 9B is a general-purpose LoRA fine-tune of
unsloth/Qwen3.5-9B, trained on 17,693 curated examples spanning code review, debugging, reasoning, and creative writing. Designed for fast, capable local inference on consumer GPUs.
Intended Use
Frank 9B is built for local deployment. It targets the sweet spot between size and capability, running comfortably on an RTX 3090 at IQ4_XS quantization (4.9 GB). It handles code review, debugging, structured reasoning, and creative writing with well-formatted, concise responses.
Where the Fred models from TendieLabs specialize in Mermaid diagram generation, Frank 9B is a generalist. The goal is a capable, direct assistant with structured output quality for everyday developer and academic use.
Training
| Detail | Value |
|---|
| Base Model | unsloth/Qwen3.5-9B |
| Method | LoRA (rank 16, alpha 32) via transformers + PEFT |
| Quantization | 8-bit (bitsandbytes) during training |
| Dataset | 17,693 rows (merged multi-source) |
| Epochs | 1 |
| Steps | 2,212 |
| Batch Size | 1 (gradient accumulation 8) |
| Max Sequence Length | 768 |
| Learning Rate | 2e-5 (cosine schedule, 30 warmup steps) |
| Final Train Loss | 0.6828 |
| Training Time | ~16 hours on RTX 3090 (24 GB) |
| Hardware | NVIDIA RTX 3090, CUDA 12.8, Docker container |
LoRA target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training Data
The merged dataset combines multiple high-quality sources:
- microsoft/rStar-Coder -- Math and code reasoning from competitive programming and STEM problem sets.
- NickyNicky/Code-290k -- Multi-language code examples across 15 programming languages.
- Crownelius/GLM-5.0-25000x -- General reasoning distillation with thinking traces.
- Crownelius/Opus-4.6-Reasoning-3300x -- Claude Opus 4.6 reasoning chains for structured analytical output.
- Crownelius/Opus-4.5-Writing-Style-formatted -- Claude writing style and tone.
- Crownelius/High-Coder-Reasoning-Multi-Turn -- Multi-turn coding sessions.
- Crownelius/Creative-Writing-Sonnet4.6-Cleaned -- Creative writing from Claude Sonnet 4.6.
Eval Results
Compared against the unmodified unsloth/Qwen3.5-9B base model on 5 prompts. Both models were loaded in 8-bit quantization with identical generation parameters (temperature 0.7, top_p 0.9, max 512 tokens). Frank 9B won or tied on every task.
| Task | Base Model | Frank 9B | Result |
|---|
Code Review (Python merge_counts) | Catches mutation bug, suggests Counter | Same catches + structured markdown headers, more thorough analysis | Frank wins |
| Debugging (FastAPI 422 error) | Correct diagnosis, two fix options | Clearer formatting, adds runnable verification code | Frank wins |
| Explain (Mutex vs Semaphore) | Good comparison table + explanation | Better structure, includes concrete Python code example | Frank wins |
| General Reasoning (sheep riddle) | Correct but verbose (158 tokens) | Correct and concise (27 tokens) | Frank wins |
| Creative Code (longest increasing subsequence) | Correct O(n^2) DP implementation | Correct, clean implementation | Tie |
Key improvements over the base model:
- Better response formatting (markdown headers, tables, structured sections)
- More concise where appropriate (no unnecessary verbosity)
- Includes practical, runnable code examples in explanations
- Same or better correctness across all evaluated tasks
Quantization
An IQ4_XS GGUF quantization is available at 4.9 GB (4.66 bits per weight). IQ4_XS is recommended for MoE and dense models alike as it uses importance-weighted quantization.
| Format | Size | BPW |
|---|
| F16 (safetensors) | 17 GB | 16.0 |
| IQ4_XS (GGUF) | 4.9 GB | 4.66 |
Capabilities
- Code review with structured, actionable feedback
- Debugging diagnosis with working fix examples
- Multi-language code generation (Python, JavaScript, Java, Rust, Go, and more)
- Mathematical reasoning and step-by-step problem solving
- Concise, well-formatted responses with markdown structure
- Fast local inference on consumer GPUs
Limitations
Frank 9B is optimized for local consumer hardware use. It is not intended for safety-critical or high-stakes applications. Performance on domains outside the training mix (e.g., medical, legal) has not been evaluated.
Developer
