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:

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

TendieLabs

Downloads last month
399
Safetensors
Model size
9B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for TendieLabs/Frank-9B

Finetuned
Qwen/Qwen3.5-9B
Adapter
(83)
this model
Adapters
3 models

Datasets used to train TendieLabs/Frank-9B