VladHong/Qwen3-4B-Instruct-Alpha

Qwen3-4B Instruct Alpha

Finetuned from unsloth/Qwen3-4B-Instruct-2507 using QLoRA + Unsloth. Finance-domain specialized — only finance and investing-related examples were retained for training. <think> blocks stripped from all assistant turns.

Key Differences from NoThink-V2

Unlike NoThink-V2 which trained on the full general-purpose dataset, Alpha applies a finance/investing keyword filter before training, resulting in a smaller but domain-focused dataset. It also incorporates a custom first-party dataset (VladHong/Alpha-Instruct) alongside the TeichAI sources.

Training Data

Dataset	Raw	After Finance Filter
TeichAI/gemini-3-pro-preview-high-reasoning-250x	248	120
TeichAI/gemini-3-pro-preview-high-reasoning-1000x	1,018	671
TeichAI/claude-4.5-opus-high-reasoning-250x	250	159
TeichAI/claude-sonnet-4.5-high-reasoning-250x	247	91
TeichAI/gpt-5.2-high-reasoning-250x	249	242
VladHong/Alpha-Instruct	336	265
Total	2,348	1,548

~1,425 examples after MinHash deduplication (threshold 0.8). Finance filter covers equities, bonds, funds, crypto, macro indicators, derivatives, retirement accounts, and more.

Training Details

Parameter	Value
Method	QLoRA (4-bit NF4) + Unsloth
LoRA rank	16
LoRA alpha	16
Epochs	1
Steps	179
Batch size	2 per device × 4 gradient accumulation = 8 effective
Learning rate	1e-4 (cosine schedule)
Max seq length	1024
Optimizer	AdamW 8-bit
Hardware	Kaggle Tesla T4 (14.56 GB VRAM)
Training time	~70.6 minutes
Trainable params	33M / 4.05B (0.81%)
Peak VRAM	5.47 GB (1.66 GB for LoRA)

Training used train_on_responses_only — loss computed on assistant completions only.

Files

• *.gguf — IQ4_XS quantized, ready for LM Studio / Ollama / llama.cpp
• lora-adapter/ — Raw LoRA weights for merging with the base model

Usage (Ollama)

ollama run VladHong/Qwen3-4B-Instruct-Alpha

License Note

Base model is Apache 2.0. Training data includes AI-generated content and a custom first-party dataset — review upstream dataset terms before commercial use.