Frox-Nano_1

Frox Morph Nano 1.1 — Trained Checkpoint

This is the nano tier of Frox AI Morph 1.1, trained for 400 pretrain + 400 SFT steps.

Files

• model.pt — merged model weights (base + LoRA, full model)
• config.json — MorphTextConfig (matches config/model_config.py)
• tokenizer/ — trained 64K BPE tokenizer (HF format)
• model/, tokenizer/, utils/, config/, inference/, scripts/ — source to load & run

Load & generate

import sys; sys.path.insert(0, ".")
import torch
from model.architecture.morph_model import MorphForCausalLM
from tokenizer.morph_tokenizer import build_morph_tokenizer

tok = build_morph_tokenizer(tokenizer_path="tokenizer")
m = MorphForCausalLM.from_saved(".", device="cpu").to("cuda")
m.eval()
ids = torch.tensor([tok.encode("Hello!", add_special_tokens=True)], device="cuda")
out = m.generate(ids, max_new_tokens=64, do_sample=True, temperature=0.7)
print(tok.decode(out[0].tolist(), skip_special_tokens=True))

Note

400 steps is a short run — the model is functional but not yet coherent. Train longer (pretrain ~50k+, SFT ~10k+) for production quality.