inspectus 0.0.5

Analytics for LLMs

Inspectus

Inspectus is a versatile visualization tool for large language models. It runs smoothly in Jupyter notebooks via an easy-to-use Python API. Inspectus provides multiple views, offering diverse insights into language model behaviors.

Preview

Click a token to select it and deselect others. Clicking again will select all again. To change the state of only one token, do shift+click

Components

Attention Matrix: Visualizes the attention scores between tokens, highlighting how each token focuses on others during processing.

Query Token Heatmap: Shows the sum of attention scores between each query and selected key tokens

Key Token Heatmap: Shows the sum of attention scores between each key and selected query tokens

Dimension Heatmap: Shows the sum of attention scores for each item in dimensions (Layers and Heads) normalized over the dimension.

Getting Started

Installation

pip install inspectus

Usage

Import the library

from inspectus import attention

Simple usage

# attn: Attention map; a 2-4D tensor or attention maps from Huggingface transformers
attention(attn, tokens)

For different query and key tokens

attention(attns, query_tokens, key_tokens)

For detailed API documentation, please refer to the official documentation - wip.

Tutorials

Huggingface model

from transformers import AutoTokenizer, GPT2LMHeadModel, AutoConfig
import torch
from inspectus import attention

# Initialize the tokenizer and model
context_length = 128
tokenizer = AutoTokenizer.from_pretrained("huggingface-course/code-search-net-tokenizer")

config = AutoConfig.from_pretrained(
    "gpt2",
    vocab_size=len(tokenizer),
    n_ctx=context_length,
    bos_token_id=tokenizer.bos_token_id,
    eos_token_id=tokenizer.eos_token_id,
)

model = GPT2LMHeadModel(config)

# Tokenize the input text
text= 'The quick brown fox jumps over the lazy dog'
tokenized = tokenizer(
    text,
    return_tensors='pt',
    return_offsets_mapping=True
)
input_ids = tokenized['input_ids']

tokens = [text[s: e] for s, e in tokenized['offset_mapping'][0]]

with torch.no_grad():
    res = model(input_ids=input_ids.to(model.device), output_attentions=True)

# Visualize the attention maps using the Inspectus library
attention(res['attentions'], tokens)

Check out the notebook here: Huggingface Tutorial

Custom attention map

import numpy as np
from inspectus import attention

# 2D attention representing attention values between Query and Key tokens
attn = np.random.rand(3, 3)

# Visualize the attention values using the Inspectus library
# The first argument is the attention matrix
# The second argument is the list of query tokens
# The third argument is the list of key tokens
attention(arr, ['a', 'b', 'c'], ['d', 'e', 'f'])

Check out the notebook here: Custom attention map tutorial