marvy-14B-lora

LoRA adapter for marvy-14B — the first open model for the full ServiceNow delivery lifecycle. Compose on top of Qwen2.5-14B-Instruct.

This is the adapter-only release (~175 MB). Apply it on Qwen/Qwen2.5-14B-Instruct to specialize the base for end-to-end ServiceNow delivery work. For ready-to-run weights see the merged model MainStack/marvy-14B or the quantized MainStack/marvy-14B-GGUF.

Released under Apache-2.0. Built with Qwen — see NOTICE.

📖 Full usage (all runtimes + OpenCode wiring): USAGE.md · Validate it works: VALIDATION.md

What it does

Fine-tunes the base for business analysis, requirements, stakeholder mapping, systems inventory, Solution Design Documents, user stories with acceptance criteria, implementation planning, test-case generation, validation/critique, and end-to-end delivery chains (story → implementation → test).

Usage

MLX (Apple Silicon)

pip install mlx-lm
python -m mlx_lm generate \
  --model Qwen/Qwen2.5-14B-Instruct \
  --adapter-path .  \
  --system-prompt "You are a senior ServiceNow delivery consultant..." \
  --prompt "Write a user story with acceptance criteria for P1 SLA escalation." \
  --max-tokens 1024 --temp 0.4

PEFT (Transformers)

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-14B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "MainStack/marvy-14B-lora")

Note: the adapter was trained with MLX-LM. The MLX adapter_config.json / adapters.safetensors are included. A PEFT-format conversion is provided for Transformers users where available; otherwise prefer the MLX path or the merged model.

Training summary

Setting	Value
Method	LoRA SFT (rank 32, scale 20, dropout 0.0)
Target keys	q/k/v/o_proj, gate/up/down_proj (top 16 layers)
Max seq length	8,192
Effective batch	16 (batch 1 × grad-accum 16)
Best checkpoint	iter 150 (best validation loss)
Framework	MLX-LM 0.31.3 on Apple Silicon

See the merged model card for full dataset, evaluation, and limitations.