31 lines
834 B
Python
31 lines
834 B
Python
import torch
|
|
from model.brain import model, tokenizer, DEVICE
|
|
from model.trainer import train_on_message
|
|
from model.dynamic_expand import expand_model_if_needed
|
|
|
|
|
|
def simulate_conversation():
|
|
expand_model_if_needed()
|
|
|
|
model.eval()
|
|
|
|
max_token_id = model.head.out_features - 1
|
|
if max_token_id < 1:
|
|
return # Safeguard if model is still too small
|
|
|
|
seed = torch.randint(0, max_token_id + 1, (1, 5), device=DEVICE)
|
|
seed = seed[:, -128:] # Clamp sequence length
|
|
|
|
output = model(seed)
|
|
|
|
preds = torch.argmax(output, dim=-1).squeeze().tolist()
|
|
if isinstance(preds, int):
|
|
preds = [preds]
|
|
|
|
# 🛡 Clamp predictions too
|
|
preds = [min(max(p, 0), max_token_id) for p in preds]
|
|
|
|
text = tokenizer.detokenize(preds)
|
|
if text and len(text.split()) >= 3:
|
|
train_on_message(text)
|