TokenProbe visualizes how language models think in real-time by showing:
- Token-by-token predictions
- Probability distributions for each prediction
- Alternative tokens the model considered
Explore how language models process and generate text, token by token
TokenProbe visualizes how language models think in real-time by showing:
Customize your generation with these parameters:
Test how the model converts natural language questions about movies into structured queries:
What movies won Best Picture Oscar in the 1990s?
{'query_type': 'award_search', 'award': 'Oscar', 'category': 'Best Picture', 'time_period': '1990s'}
Show me films directed by Christopher Nolan after 2010
{'query_type': 'director_search', 'director': 'Christopher Nolan', 'year_from': 2010}
Tell me about movies featuring both Tom Hanks and Meg Ryan
Explore how the model handles cricket statistics queries:
What was MS Dhoni's batting average in IPL 2023?
{'query_type': 'player_stats', 'player': 'MS Dhoni', 'stat': 'batting_average', 'tournament': 'IPL', 'year': 2023}
List all centuries scored in the 2023 World Cup
{'query_type': 'tournament_stats', 'stat': 'centuries', 'tournament':
'World Cup', 'year': 2023}
Who was the highest scorer in BBL 2018?
Watch the model learn to generate code based on natural language descriptions:
Write a Python function that calculates the factorial of a number
def factorial(n):
if n == 0:
return 1
return n * factorial(n-1)
Write a function that checks if a string is a palindrome
def is_palindrome(s):
return s == s[::-1]
Write a function that finds the maximum element in a list
Observe how the model's confidence increases with each example as it learns the pattern of translating requirements into code.
Test how the model learns to convert natural language questions into SQL queries:
Find all customers who made purchases over $1000
SELECT DISTINCT customer_name
FROM orders
WHERE order_amount > 1000;
Show me the total sales by category for last quarter
SELECT category, SUM(sales) as total_sales
FROM sales
WHERE quarter = 'Q4'
GROUP BY category;
List employees who haven't made a sale this month
Watch the model learn the pattern and predict the appropriate SQL structure for the final query.
This example demonstrates how the model exhibits Bayesian learning even in general text completion, not just in-context learning scenarios:
In an unprecedented move that surprised industry analysts, tech giant MicroDyne appointed Sarah Chen Zhao as their new CEO last quarter. The company's board cited her extensive experience in AI and quantum computing as key factors in their decision. Market observers noted that under previous leadership, MicroDyne had struggled to keep pace with rapid technological changes. When asked about the company's strategic direction and the recent slump in its stock price,
[Watch how the model learns to use the full name and position, continuing in a natural news style]
Here's another business news excerpt:
The growing concerns about artificial intelligence regulation have prompted major tech companies to form a new industry alliance. Market analysts were particularly interested in MicroDyne's stance, given their recent technological advances. When reached for comment,
Observe how the model learns and maintains consistency with the name "Sarah Chen Zhao" and her position as CEO, showing high confidence in using this information in the completion.
Here's another example showing Bayesian learning in scientific writing:
A groundbreaking study by Dr. Aisha Patel-Rodriguez at the Quantum Materials Institute has demonstrated a novel approach to room-temperature superconductivity. Using a previously unexplored combination of rare earth elements, specifically dysprosium and ytterbium in a hexagonal lattice structure, the team achieved sustained superconductivity at 294K. The implications for energy transmission efficiency are profound, though scaling remains a challenge. When asked about potential industrial applications,
[Notice how the model learns the researcher's name and institutional context]
In related developments from the field of materials science, recent advances in quantum computing have opened new avenues for simulating complex molecular structures. The theoretical predictions align closely with experimental results, particularly in the domain of high-temperature superconductivity. Regarding these computational findings,
Watch how the model demonstrates Bayesian learning by: