ramankrishna10/npc-agentic-7b-v3-lora

NPC Agentic 7B — LoRA adapter

LoRA adapter for NPC Agentic 7B. Apply on top of Qwen/Qwen2.5-7B-Instruct (or load the merged FP16 model from the sibling repo if you want a ready-to-run artifact).

See ramankrishna10/npc-agentic-7b for the full training recipe, eval numbers, and known limitations.

Training config

• rank = 64, alpha = 128, dropout = 0.05
• target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• trained with QLoRA 4-bit base, bf16 adapters, Unsloth + TRL 0.24
• 11,410 steps, 2 epochs, ~96 GPU-hours on A40

Use

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

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct", torch_dtype=torch.float16, device_map="auto",
)
model = PeftModel.from_pretrained(base, "ramankrishna10/npc-agentic-7b-v3-lora")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

# Optional: bake the adapter in for faster inference
# model = model.merge_and_unload()

Adapter size

~616 MB safetensors (161.5M trainable params).

---

Built by Bottensor.

Citation

If you use NPC Agentic 7B in your work, please cite:

@misc{bachu2026npcagentic7b,
  title        = {NPC Agentic 7B: A Single-GPU QLoRA Recipe for a Laptop-Scale Conversational Model},
  author       = {Bachu, Rama Krishna},
  year         = {2026},
  month        = may,
  publisher    = {Zenodo},
  version      = {v1},
  doi          = {10.5281/zenodo.19954103},
  url          = {https://doi.org/10.5281/zenodo.19954103},
  note         = {Preprint}
}

Paper: https://doi.org/10.5281/zenodo.19954103