hai3.1-checkpoint-0001

CURRENTLY IN TRAINING :)

Currently, only the LLM section of this model is fully ready.

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
import torch

# Load model and tokenizer
model_name = "Abhaykoul/hai3.1-pretrainedv3"

# Set device to CUDA if available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, torch_dtype="auto")
model.to(device)
print(model)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)

# Message role format for chat
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": """hlo"""},
]

# Apply chat template to format prompt
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Tokenize input and move to device
inputs = tokenizer(prompt, return_tensors="pt")
inputs = {k: v.to(device) for k, v in inputs.items()}

# Set up text streamer for live output
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

# Generate text with streaming
model.generate(
    **inputs,
    max_new_tokens=4089,
    temperature=0.7,
    top_p=0.9,
    do_sample=True,
    streamer=streamer
)

Classfication section undertraining

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

ckpt = "Abhaykoul/hai3.1-pretrainedv3"
device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(ckpt, trust_remote_code=True).to(device).eval()
tok = AutoTokenizer.from_pretrained(ckpt, trust_remote_code=True)
if tok.pad_token is None:
    tok.pad_token = tok.eos_token

text = "I am thrilled about my new job!"
enc = tok([text], padding=True, truncation=True, max_length=2048, return_tensors="pt")
enc = {k: v.to(device) for k, v in enc.items()}

with torch.no_grad():
    out = model(input_ids=enc["input_ids"], attention_mask=enc.get("attention_mask"), output_hidden_states=True, return_dict=True, use_cache=False)
    last = out.hidden_states[-1]
    idx = (enc["attention_mask"].sum(dim=1) - 1).clamp(min=0)
    pooled = last[torch.arange(last.size(0)), idx]
    logits = model.structured_lm_head(pooled)
    pred_id = logits.argmax(dim=-1).item()

print("Predicted class id:", pred_id)
# Map id -> label using your dataset’s label list, e.g.:
id2label = ["sadness","joy","love","anger","fear","surprise"]  # dair-ai/emotion
print("Predicted label:", id2label[pred_id] if pred_id < len(id2label) else "unknown")

TTS layers in training

NOTE: we have used qwen2 tokenizer in it

This model contains layers from our diffrent models To aline layers we have done post training after merging layers