cminst/Llama-Nemotron-8B-templatefixes

Llama-3.1-Nemotron-Nano-8B-v1

Note: chat template forces reasoning to be on via the system prompt! Any additional system prompt will throw an error.

Example:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "cminst/Llama-Nemotron-8B-templatefixes"

# Load tokenizer + override chat template
tokenizer = AutoTokenizer.from_pretrained(model_id)

# ---- Test conversation ----
messages = [
    {"role": "user", "content": "Solve x*(sin(x)+2)=0"}
]

# Apply template
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    return_tensors="pt",
    add_generation_prompt=True
)
print("START")
print(inputs,end="")
print("END")

gives:

START
<|begin_of_text|><|start_header_id|>system<|end_header_id|>

detailed thinking on<|eot_id|><|start_header_id|>user<|end_header_id|>

Solve x*(sin(x)+2)=0<|eot_id|><|start_header_id|>assistant<|end_header_id|>

END