BlackSwan v5

BlackSwan is a fine-tuned AI accountability partner built for The Underground Circle — a productivity app for dev teams.

Model Details

  • Base model: Qwen/Qwen3.5-4B
  • Fine-tuning: LoRA (rank 16, alpha 16) on 40K+ training examples
  • Training: 2500 iterations, final loss 6.3
  • Format: safetensors (bfloat16)
  • License: Apache 2.0

Personality

BlackSwan has quiet confidence — knowledgeable but never arrogant. Direct. No fluff, no corporate speak. Gives real feedback. Knows productivity, design, UI/UX, coding, architecture, and general knowledge.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("cswan801/BlackSwan-v5")
tokenizer = AutoTokenizer.from_pretrained("cswan801/BlackSwan-v5")

Training Data

  • Code instructions (CodeAlpaca, Evol-Instruct-Code)
  • General knowledge (OpenHermes 2.5, SlimOrca)
  • Multi-turn conversation (Capybara, UltraChat)
  • Math reasoning (GSM8K, MathInstruct)
  • App-specific data (terminal commands, agent activity, tasks, check-ins)
Downloads last month
690
Safetensors
Model size
5B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for cswan801/BlackSwan-v5

Finetuned
Qwen/Qwen3.5-4B
Adapter
(93)
this model