ShaziLLM-1B

ShaziLLM-1B is a compact 1.08B-parameter instruction and reasoning language model designed for coding, structured instruction following, tool use, and local deployment. This repository contains the normal BF16 Transformers checkpoint.

For lower-memory CPU and browser inference, use ShaziLLM-1B-ONNX.

Quick start

pip install -U transformers torch accelerate

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Sharjeelbaig/ShaziLLM-1B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Write a Python function to merge two sorted lists."}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
    )

new_tokens = output[0, inputs.input_ids.shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Model details

The configured context limit is not a practical memory recommendation. Start with a much shorter context on low-memory hardware and increase it only after measuring KV-cache usage.

ShaziLLM alignment

The release received a focused identity-alignment and capability-retention pass before export. A rank-8 LoRA adapter targeting the query and value projections was trained for four epochs at a 1e-4 learning rate on 48 short examples, with 1,032,192 trainable parameters. The adapter was merged into the BF16 checkpoint. Validation probes covered model identity, arithmetic, coding, and factual recall.

Capabilities

• Code generation and debugging
• Instruction following and structured responses
• Reasoning-style responses
• XML-formatted tool calls through the included chat template
• English and Chinese text generation

Low-hardware deployment

The BF16 checkpoint requires roughly 2.1 GB for weights alone, plus runtime and KV-cache memory. The separate ONNX INT8 release reduces weight storage and CPU inference memory substantially. Limit context and generated-token counts on constrained devices.

Limitations

ShaziLLM-1B can hallucinate, produce incorrect code, and expose reasoning-like text. Do not treat its output as professional, security, legal, medical, or financial advice. Validate generated code before execution and isolate tool access behind deterministic authorization checks.

License and attribution

Released under Apache License 2.0; see LICENSE. ShaziLLM-1B is a fine-tuned derivative of OpenBMB/MiniCPM5-1B. The ShaziLLM name identifies this release and does not imply authorship of the upstream base architecture.