ArgusForge/qwen2.5-14b-dolly-lora

Qwen2.5-14B-Instruct, dolly LoRA (demo)

A small LoRA adapter for Qwen2.5-14B-Instruct, trained on databricks-dolly-15k. This is a demonstration artifact. It exists to show a full open-model loop end to end (fine-tune, merge, quantize, serve), not to be a state-of-the-art model.

What it does

Trained on dolly's concise, human-written responses, the adapter shifts the base model away from its long multi-section default toward shorter, more direct answers.

Training

• Base: Qwen/Qwen2.5-14B-Instruct
• Data: databricks/databricks-dolly-15k (first 3000 examples)
• Method: LoRA (r=16, alpha=32, dropout=0.05) on attention and MLP projections, bf16
• Steps: 200, effective batch 16, lr 2e-4, cosine schedule
• Trainable params: 68.8M (0.46 percent of 14.8B)
• Final training loss: 1.21 (down from 1.98)
• Hardware: one H100 SXM 80GB

Before and after (same prompt, greedy decoding)

Prompt: Explain why the sky is blue to a ten year old.

• Base: a long four-point numbered explanation with section headers.
• Base plus this adapter: a single concise paragraph in dolly's style.

Use

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen2.5-14B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "ArgusForge/qwen2.5-14b-dolly-lora")

Related

• Quantized GGUF of the merged model: ArgusForge/qwen2.5-14b-dolly-q4_K_M-GGUF
• Benchmark and training code (the full open-model loop): https://github.com/argus-forge/open-model-ops