JupiterJil/cyberspace-qwen25-7b-lora

Cyberspace Presales Co-Pilot — Qwen 2.5 7B LoRA Adapter

LoRA adapter for Cyberspace Limited's Vertical AI Presales Co-Pilot.

Training Summary

• Base model: Qwen/Qwen2.5-7B-Instruct
• Method: LoRA (QLoRA 4-bit) via Unsloth
• Dataset: 132 Cyberspace proposals (ChatML format, zero template bleed)
• Final loss: 1.4795 (target: 1.2–1.6 ✅)
• Training time: ~4.9 min on A100 80GB

LoRA Config

• Rank: 16
• Alpha: 16
• Dropout: 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Epochs: 3
• LR: 2e-4 (cosine scheduler)
• train_on_responses_only: True (ChatML markers)

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "JupiterJil/cyberspace-qwen25-7b-lora")

Recommended Inference Settings

• Temperature: 0.4
• Top-p: 0.9
• Repeat penalty: 1.1
• Use detailed category-specific system prompts

Production Artifact

For inference, use the GGUF Q4_K_M version: JupiterJil/cyberspace-qwen25-7b-gguf