Trained on 7,106 instruction pairs extracted from 76 textbooks spanning algorithms, system design, networking, databases, compilers, and programming languages (Python, Go, Rust, JavaScript, C++). Answers are grounded in real engineering source material, not internet-scale guesswork.

Model Details

Property	Value
Base Model	Qwen2.5-14B-Instruct
Fine-tuning	QLoRA (4-bit NF4), LoRA rank 8, alpha 16, dropout 0.05
Training Data	7,106 high-quality QA pairs (balanced, deduplicated, task-validated)
Task Types	15 modes: code explanation, debugging, review, generation, complexity analysis, testing, modernization, full implementation, code completion, production scenarios, cross-source synthesis, diagram generation, prose, math, quizzes
Data Mix	40% code, 20% debug, 25% design, 15% docs/Q&A
Epochs	2
Effective Batch Size	32 (micro-batch 1 × 32 accum)
Learning Rate	1e-4
Warmup	5%
GPU	A100-40GB
Training Time	164 min
Language Coverage	Python, Go, Rust, JavaScript, TypeScript, Java, C, C++, SQL

Evaluation (50 held-out QA pairs)

Metric	Value
Average QA Score	3.3/5
Hallucination Rate	~2%
Strongest Domains	Algorithms, Databases, Linux, Web APIs
Weakest Domains	Interview prep, abstract design principles

Scored by human audit against ground-truth answers generated from the same source textbooks.

Capabilities

Explain CS concepts from 70+ textbooks with source attribution
Write production-grade code in Python, Go, Rust, JS, TS, Java, C with modern syntax (2024 editions)
Debug and review code — find subtle bugs, suggest improvements with rationale
Analyze algorithm complexity — trace loops, recursion, data structures
Synthesize across books — when sources disagree, attribute each position
Generate Mermaid diagrams — architecture, sequences, class hierarchies
Multi-turn tutoring — conversational teaching with follow-ups

Files

File	Size	Use
`fwen-14b-q4_k_m-v1.gguf`	8 GB	Production serving (Q4_K_M quantization)
`fwen-14b-q8_0-v1.gguf`	14 GB	Benchmark evaluation (Q8_0 quantization)

Usage

Ollama

bash

ollama create fwen:14b -f Modelfile
ollama run fwen:14b

llama.cpp

./llama-cli -m fwen-14b-q4_k_m-v1.gguf -p "Explain Rust ownership" -n 512

HF Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("friehub/fwen-14b-v1")

Limitations

Trained on 7K pairs — smaller than typical fine-tune datasets. Expect adequate but not expert-level depth.
Rust/Go underrepresented in training data (v2 will address this).
Source attribution is sometimes vague ("the text emphasizes" vs "Chapter 4 of Code Simplicity states").

Training Pipeline

PDF ingestion → domain classification → code/math/prose unit extraction (212,616 units)
15 task-type instruction pairs generated by DeepSeek V4 Pro/Flash
Quality filter: length, dedup, relevance (cosine similarity), task validators
Language balance: capped at 25% Python code pairs
QLoRA fine-tune with Unsloth on A100-40GB

Citation

@model{fwen-14b-v1,
  author = {Friehub},
  title = {Fwen-14B: A Fine-Tuned Software Engineering Tutor},
  version = {v1},
  year = 2026,
  url = {https://huggingface.co/friehub/fwen-14b-v1}
}

Friehub/fwen-14b-v1

Fwen-14B-v1