Datasets:
id stringlengths 6 15 | prompt stringlengths 12 3.8k | category stringclasses 9
values | tier stringclasses 3
values | reference_output stringclasses 1
value | reference_model stringclasses 2
values | source stringclasses 15
values |
|---|---|---|---|---|---|---|
qa_001 | What is the capital of Argentina? | qa | low | claude-opus-4-7 | null | |
qa_003 | What does HTTP stand for? | qa | low | claude-opus-4-7 | null | |
qa_004 | Translate 'Where is the train station?' to French. | qa | low | claude-opus-4-7 | null | |
qa_005 | What is the chemical symbol for tungsten? | qa | low | claude-opus-4-7 | null | |
qa_006 | How many time zones does Russia have? | qa | low | claude-opus-4-7 | null | |
qa_007 | Convert 75°F to Celsius. | qa | low | claude-opus-4-7 | null | |
qa_008 | What is the difference between viral and bacterial infections in one sentence? | qa | low | claude-opus-4-7 | null | |
qa_009 | Define 'concurrency' in computer science. | qa | low | claude-opus-4-7 | null | |
qa_010 | What is the boiling point of nitrogen? | qa | low | claude-opus-4-7 | null | |
sum_001 | Summarize this paragraph in one sentence: The remote work shift accelerated in 2020 has led to permanent changes in commercial real estate, with vacancy rates in major US cities staying elevated through 2025. Companies that adopted hybrid models report higher retention but lower spontaneous collaboration. | summarization | low | claude-opus-4-7 | null | |
sum_002 | Give me a TL;DR of this customer review: 'I bought this espresso machine three months ago and at first it was great — pulled good shots, easy to use. But after about six weeks the steam wand started clogging and now it barely works. Customer service was unhelpful when I called. Will not buy again.' | summarization | low | claude-opus-4-7 | null | |
sum_003 | In three bullet points, summarize the key arguments for and against a 4-day work week based on common reporting: productivity holds or improves, but coordination across time zones gets harder; employee satisfaction goes up but customer service hours need rethinking; recruiting becomes easier but legacy contract terms c... | summarization | medium | claude-opus-4-7 | null | |
sum_004 | Summarize this meeting transcript snippet: 'Alex: We need to ship the new search by Q3. Sam: That's tight — we'd have to cut the autocomplete work. Alex: Fine, push autocomplete to Q4. Sam: Then we need approval to add Priya to the team. Alex: I'll talk to leadership.' | summarization | medium | claude-opus-4-7 | null | |
sum_005 | Condense this product feedback into 3 themes: 'I love how fast the app is — switching between projects feels instant. The notifications are too aggressive though, I keep getting buzzed for things I don't care about. Also the search is great but I wish I could filter by date. Overall, much better than what we used befor... | summarization | medium | claude-opus-4-7 | null | |
sum_006 | Brief summary in two sentences of this article preview: 'A new study from Stanford analyzed 50,000 GitHub repositories and found that LLM-assisted commits are 23% more likely to introduce subtle bugs but 40% faster to ship. The data suggests current code review practices are insufficient to catch the new bug patterns.' | summarization | medium | claude-opus-4-7 | null | |
sum_007 | TL;DR: 'After 18 months of work, our platform team migrated 200+ services from on-prem to AWS. Initial estimates put the cost at $2.3M; final tally was $4.1M, largely from unforeseen networking and data transfer charges. Performance improved 30% but we'd recommend others to over-budget on egress.' | summarization | medium | claude-opus-4-7 | null | |
sum_008 | Summarize this support thread's resolution: 'Customer: My API key isn't working since yesterday. Support: Can you check if it expired? Customer: Yes, it expired Wednesday. Support: Generate a new one in settings → API keys. Customer: Done — works now, thanks.' | summarization | low | claude-opus-4-7 | null | |
sum_009 | Three-sentence executive summary of a hypothetical retrospective: A team launched a new pricing tier; conversions exceeded forecast by 18%, but churn at month 3 was 2× expected. Root cause was unclear messaging about included features. Recommendation: add comparison table to checkout flow and follow-up email at day 75. | summarization | medium | claude-opus-4-7 | null | |
sum_010 | Condense into one bullet list: 'The Q1 outage affected 12% of customers for 4 hours. Cause was a misconfigured deploy that broke session affinity in our load balancer. Detection was delayed because our synthetic monitor was hitting cached endpoints. Postmortem identified five fixes: pre-deploy config validation, additi... | summarization | medium | claude-opus-4-7 | null | |
ext_001 | Extract all email addresses from: 'Reach out to alice@acme.com or bob.smith@acme.com for procurement, and copy ceo@acme.com on contracts over $50K.' | extraction | low | claude-opus-4-7 | null | |
ext_002 | From this paragraph, extract: company names, dollar amounts, and dates. 'Acme Corp raised $25M Series B from Sequoia on March 15, 2025, following their $5M seed from a16z in 2023.' | extraction | medium | claude-opus-4-7 | null | |
ext_003 | List all the file paths mentioned in this stack trace: 'TypeError in /app/services/billing.py:142, called from /app/routes/orders.py:88 (request handler), originally from /app/middleware/auth.py:23.' | extraction | medium | claude-opus-4-7 | null | |
ext_004 | Extract product names and prices from: 'The Pro Max ships for $1,299. The Pro Lite is $899. The standard model starts at $649 with monthly financing at $54/mo.' | extraction | low | claude-opus-4-7 | null | |
ext_005 | From this customer support ticket, identify: issue type, urgency, and any feature names mentioned. 'I can't access the billing dashboard since the morning. Tried clearing cache. Need to update our subscription to Team Plus before our renewal Friday — urgent.' | extraction | medium | claude-opus-4-7 | null | |
ext_006 | List all the medications and dosages in this note: 'Patient is on metformin 500mg twice daily, lisinopril 10mg in the morning, and atorvastatin 20mg at bedtime.' | extraction | medium | claude-opus-4-7 | null | |
ext_007 | Extract URLs and their stated purposes from: 'Documentation is at https://docs.example.com, the API reference at https://api.example.com/v2, status page at https://status.example.com, and support ticketing at https://support.example.com/new.' | extraction | low | claude-opus-4-7 | null | |
ext_008 | From the following job posting, extract: required years of experience, programming languages, and benefits. 'We're hiring a Senior Engineer (5+ yrs) fluent in Python and Go, with strong Postgres skills. Benefits: full health/dental, unlimited PTO, equity, $4K/yr learning stipend.' | extraction | medium | claude-opus-4-7 | null | |
ext_009 | Identify all phone numbers and email addresses in: 'Call Sarah at (415) 555-0143 or email sarah.j@startup.io. For emergencies, dial 415.555.9999.' | extraction | low | claude-opus-4-7 | null | |
ext_010 | Extract the customer's order history from: 'Order #4523 placed Jan 12 for $89.99 (delivered), Order #4892 placed Feb 28 for $234.50 (in transit), Order #5101 placed Mar 18 for $52.00 (returned).' | extraction | medium | claude-opus-4-7 | null | |
code_001 | Write a Python function that takes a list of integers and returns the second largest value. | coding | low | claude-opus-4-7 | null | |
code_002 | Write a JavaScript function `debounce(fn, ms)` that returns a debounced version of fn. | coding | medium | claude-opus-4-7 | null | |
code_003 | Convert this SQL query to a Python pandas equivalent: SELECT customer_id, SUM(amount) FROM orders WHERE created_at > '2025-01-01' GROUP BY customer_id HAVING SUM(amount) > 1000 ORDER BY SUM(amount) DESC LIMIT 10; | coding | medium | claude-opus-4-7 | null | |
code_004 | Write a Python decorator that retries the wrapped function up to N times on exception, with exponential backoff. | coding | medium | claude-opus-4-7 | null | |
code_005 | Review this Python function and identify any bugs:
def get_user_avatars(user_ids):
avatars = {}
for user_id in user_ids:
avatars[user_id] = fetch_avatar(user_id)
return avatars
(Assume fetch_avatar() can raise NotFoundError.) | coding | medium | claude-opus-4-7 | null | |
code_006 | Implement a function in Go that reads a JSON file with this structure ({users: [{id, name, email}]}) and returns the emails of users whose IDs are in a given list. | coding | medium | claude-opus-4-7 | null | |
code_007 | Refactor this Python code to be more idiomatic:
def process(data):
result = []
for item in data:
if item['active'] == True:
if item['count'] > 10:
result.append(item['name'])
return result | coding | low | claude-opus-4-7 | null | |
code_008 | Write a complete React component (functional, hooks) that fetches data from /api/users on mount, shows a loading state, handles errors, and renders a list of users. | coding | medium | claude-opus-4-7 | null | |
code_009 | Implement a thread-safe LRU cache in Python without using functools.lru_cache. Include tests for capacity, eviction order, and concurrent access. | coding | high | claude-opus-4-7 | null | |
code_010 | Diagnose this stack trace and propose a fix:
TypeError: NoneType object is not subscriptable
File 'app/services/orders.py', line 142, in get_order_summary
customer_name = order['customer']['name']
File 'app/routes/orders.py', line 88, in handle_get
return get_order_summary(order_id) | coding | medium | claude-opus-4-7 | null | |
code_011 | Design and implement a Python class for a thread-safe job queue that supports: enqueue, dequeue, peek, size, and a worker decorator. Include reasoning for synchronization choices. | coding | high | claude-opus-4-7 | null | |
code_012 | Write a complete Postgres migration script (idempotent) that adds a `deleted_at` timestamp column to a `users` table of ~50M rows without blocking writes. | coding | high | claude-opus-4-7 | null | |
code_013 | Implement a debounced search input in TypeScript + React with cancellation: typing should cancel any in-flight request before starting a new one. | coding | high | claude-opus-4-7 | null | |
code_014 | Write a complete Python CLI tool using Click that takes a CSV path, a column name, and aggregates values by that column with sum/count/avg. | coding | medium | claude-opus-4-7 | null | |
code_015 | Write a SQL query that, given a `events` table with `(user_id, event_type, ts)`, computes 30-day retention curves for new signups, grouped by signup month. | coding | high | claude-opus-4-7 | null | |
rea_001 | A train leaves Station A at 60 mph. Another leaves Station B (180 miles away) at 90 mph in the opposite direction. They start at the same time. How long until they're 360 miles apart? Show your work. | reasoning | medium | claude-opus-4-7 | null | |
rea_002 | If a bat and ball cost $1.10 together, and the bat costs $1.00 more than the ball, how much does the ball cost? Walk through your reasoning. | reasoning | low | claude-opus-4-7 | null | |
rea_003 | Three people split a $30 dinner bill. Each pays $10 to the cashier. The cashier realizes the bill was actually $25 and gives the waiter $5 to return. The waiter pockets $2 and gives each person $1. Now each person paid $9, totaling $27, plus $2 the waiter has = $29. Where's the missing dollar? Walk through the logic. | reasoning | high | claude-opus-4-7 | null | |
rea_004 | If I have 4 red socks, 6 blue socks, and 8 green socks in a dark drawer, what's the minimum number of socks I need to pull to guarantee a matching pair? | reasoning | medium | claude-opus-4-7 | null | |
rea_005 | A hospital tests for a disease with 99% accuracy. The disease affects 1 in 10,000 people. If you test positive, what's the actual probability you have the disease? Show your reasoning using Bayes' theorem. | reasoning | high | claude-opus-4-7 | null | |
rea_006 | John is twice as old as Mary was when John was the age Mary is now. Mary is 30. How old is John? | reasoning | high | claude-opus-4-7 | null | |
rea_007 | Prove that the sum of the first n odd integers equals n². Use induction. | reasoning | high | claude-opus-4-7 | null | |
rea_008 | Given an array of n integers, write the most efficient algorithm to find the kth largest element. Justify the time complexity and explain the tradeoffs vs. simpler approaches. | reasoning | high | claude-opus-4-7 | null | |
rea_009 | Why does HTTPS use both asymmetric and symmetric encryption? Walk through the handshake step by step. | reasoning | medium | claude-opus-4-7 | null | |
rea_010 | I'm seeing 99th-percentile latency spike from 200ms to 2 seconds during deploys, but average stays flat. What are the likely causes, in order of probability? | reasoning | high | claude-opus-4-7 | null | |
rea_011 | Why does my Postgres query with LIMIT 10 sometimes scan a million rows? Explain in terms of the planner. | reasoning | high | claude-opus-4-7 | null | |
rea_012 | When should you use a CRDT instead of operational transformation for collaborative editing? Discuss tradeoffs. | reasoning | high | claude-opus-4-7 | null | |
rea_013 | We see memory growing 50MB/hour in a Go service. Heap profile is flat but RSS keeps climbing. What's the likely cause? | reasoning | high | claude-opus-4-7 | null | |
rea_014 | If I add a unique index on (org_id, email) to a table where org_id is heavily skewed (one tenant has 80% of rows), how will writes be affected? Explain why. | reasoning | high | claude-opus-4-7 | null | |
rea_015 | Two coins: one fair, one weighted (75% heads). I pick one at random, flip 3 times, get HHH. What's the probability I picked the weighted coin? | reasoning | high | claude-opus-4-7 | null | |
pln_001 | Draft a 6-month migration plan for moving 30 microservices from Kubernetes self-hosted to a managed platform. Include phases, risks, and rollback strategy. | planning | high | claude-opus-4-7 | null | |
pln_002 | Design a tiered alerting strategy for a SaaS company with 99.9% SLA: which signals page humans, which auto-remediate, which create tickets? | planning | high | claude-opus-4-7 | null | |
pln_003 | Propose a 30-60-90 day plan for a new engineering manager joining a team that's behind on its quarterly goals. | planning | medium | claude-opus-4-7 | null | |
pln_004 | Outline a strategy for migrating from REST to GraphQL across 50 client teams, given we can't break existing integrations. | planning | high | claude-opus-4-7 | null | |
pln_005 | Design a launch plan for a new pricing tier that includes feature gating, customer communication, A/B testing, and rollback. Include success metrics. | planning | medium | claude-opus-4-7 | null | |
pln_006 | Draft a roadmap for adding multi-region failover to a Postgres-backed service currently in one region. Cover replication strategy, DNS, and DR drills. | planning | high | claude-opus-4-7 | null | |
pln_007 | Plan a database schema migration that adds soft-deletes to a 100M-row users table without downtime. | planning | high | claude-opus-4-7 | null | |
pln_008 | Draft a one-quarter plan to reduce our AI inference costs by 40% without measurable quality regression. Include experiments, metrics, and decision gates. | planning | high | claude-opus-4-7 | null | |
pln_009 | I'm a YC founder building developer tools. Outline the next 60 days: customer discovery, MVP scope, key metrics, and de-risking experiments. | planning | medium | claude-opus-4-7 | null | |
pln_010 | Plan a customer-facing incident response process: detection, communication, postmortem, and follow-up. Include who does what at each stage. | planning | medium | claude-opus-4-7 | null | |
cre_001 | Write a haiku about a server going offline at 3 AM. | creative | low | claude-opus-4-7 | null | |
cre_002 | Draft a friendly Slack message announcing our team's new on-call rotation, including a joke that lands without being unprofessional. | creative | medium | claude-opus-4-7 | null | |
cre_003 | Write a 200-word product launch announcement for a new AI inference cost-optimization platform aimed at CTOs at SaaS companies. | creative | medium | claude-opus-4-7 | null | |
cre_004 | Write a cold email to a head of engineering at a 200-person Series C startup, offering a 30-min call about reducing their AI bill. Be specific, short, and humble. | creative | medium | claude-opus-4-7 | null | |
cre_005 | Write a one-page short story (around 300 words) about a developer who discovers their code was actually written by their future self. | creative | high | claude-opus-4-7 | null | |
cre_006 | Generate five product names for an AI inference control plane that sound serious and infrastructure-y, not whimsical. | creative | low | claude-opus-4-7 | null | |
cre_007 | Write a job description for a Senior Platform Engineer in a voice that's warm, specific to a 50-person Series B SaaS, and not full of cliches. | creative | medium | claude-opus-4-7 | null | |
cre_008 | Draft a personal mission statement for an engineer who wants to do work that compounds over decades, in their own voice, ~80 words. | creative | medium | claude-opus-4-7 | null | |
cre_009 | Write a polite but firm follow-up email to a vendor who has missed a deadline by 2 weeks and is not responding to messages. | creative | medium | claude-opus-4-7 | null | |
cre_010 | Write a tweet (280 chars) announcing the GA of a developer tool. Make it punchy, specific, and not generic AI hype. | creative | low | claude-opus-4-7 | null | |
str_001 | Return as JSON with keys 'name', 'age', 'email': My name is Alice Chen, I'm 34, and you can reach me at alice@example.com. | structured_output | low | claude-opus-4-7 | null | |
str_002 | Output a JSON array of objects {city, country, population_millions} for: London, Tokyo, Cairo, and Sydney. | structured_output | low | claude-opus-4-7 | null | |
str_003 | Return YAML representing a Kubernetes Deployment for a service called 'api', image 'myapp/api:v2', 3 replicas, exposing port 8080. | structured_output | medium | claude-opus-4-7 | null | |
str_004 | Generate a JSON schema describing an order: order_id (string), customer_id (string), items (array of {sku, quantity, price}), total (number), placed_at (ISO date). | structured_output | medium | claude-opus-4-7 | null | |
str_005 | Parse this resume excerpt into structured JSON {name, title, years_experience, skills (array), companies (array of {name, role, years})}: 'Sarah Kim - Senior Backend Engineer, 8 years experience in Python, Go, Postgres. Previously at Stripe (Backend Eng, 3 yrs) and Cloudflare (Senior Backend Eng, 5 yrs).' | structured_output | medium | claude-opus-4-7 | null | |
str_006 | Return CSV format: header line, then rows. Data: Alice/30/Manager, Bob/27/IC, Carol/45/Director. | structured_output | low | claude-opus-4-7 | null | |
str_007 | Output as Markdown table: top 5 programming languages by popularity in 2025 with columns: Language, Rank, Primary Use Case. | structured_output | low | claude-opus-4-7 | null | |
str_008 | Generate a JSON config for a CI pipeline with: lint stage, test stage (with services postgres+redis), build stage (only on main), deploy stage (manual approval). | structured_output | medium | claude-opus-4-7 | null | |
str_009 | Convert this prose to a structured JSON event log: 'At 14:23 Alice logged in. At 14:25 she viewed the dashboard. At 14:30 she exported a report.' | structured_output | medium | claude-opus-4-7 | null | |
str_010 | Return the following as JSON, with units explicit: 'The server has 32GB of RAM, 8 cores, 1TB SSD, and runs Ubuntu 24.04.' | structured_output | low | claude-opus-4-7 | null | |
lc_001 | Synthesize three perspectives on the impact of generative AI on software engineering: (1) optimistic — productivity multipliers and democratization, (2) pessimistic — bug surface area and skill atrophy, (3) skeptical — most claims are unfalsifiable. Then offer your reconciled view in one paragraph. | long_context | high | claude-opus-4-7 | null | |
lc_002 | Compare and contrast event sourcing vs. CQRS vs. outbox pattern for ensuring consistency in a multi-service order management system. Cover failure modes, observability, and operational complexity. | long_context | high | claude-opus-4-7 | null | |
lc_003 | Walk through the design tradeoffs of B-trees, LSM trees, and skip lists for a key-value store optimized for write-heavy workloads. Explain why most modern OLTP databases ended up choosing what they did. | long_context | high | claude-opus-4-7 | null | |
lc_004 | Critique the following architecture proposal and identify three risks I haven't addressed: 'We'll use Kafka for event distribution, Postgres for state, and Redis for caching. Services will publish events on state changes and consume to update their views.' | long_context | high | claude-opus-4-7 | null | |
lc_005 | Given a multi-tenant SaaS that grew from 100 to 10,000 customers in 18 months, outline how the database access patterns probably changed and what changes the team should consider now. | long_context | high | claude-opus-4-7 | null | |
lc_006 | Explain how a vector database serves nearest-neighbor queries on 100M vectors at p99 < 50ms. Cover HNSW vs. IVF tradeoffs and where the bottlenecks are at scale. | long_context | high | claude-opus-4-7 | null | |
lc_007 | Walk through how to safely roll out a breaking API change to 500 client teams, given that some are on auto-update and others are pinned. Include rollback strategy. | long_context | high | claude-opus-4-7 | null | |
lc_008 | Design a recommendation system for a marketplace with 10M items and 50M users that updates embeddings near-real-time. Cover the data pipeline, model architecture, serving infrastructure, and online evaluation. | long_context | high | claude-opus-4-7 | null | |
lc_009 | Explain why distributed transactions are hard and walk through three strategies (2PC, sagas, idempotent consumers) with their tradeoffs in a microservices order pipeline. | long_context | high | claude-opus-4-7 | null | |
lc_010 | What are the strongest counterarguments to the claim that LLMs will be net-positive for software engineering productivity over the next 5 years? Steelman each, then offer your own assessment. | long_context | high | claude-opus-4-7 | null | |
hf_0001_rea_hig | Answer this question, with reasoning where appropriate:
Typical advertising regulatory bodies suggest, for example that adverts must not: encourage _________, cause unnecessary ________ or _____, and must not cause _______ offence. | reasoning | high | reasoning |
End of preview. Expand in Data Studio
LLM Downshifting Evaluation Artifacts
This repository accompanies How Much Quality Survives LLM Downshifting? Judge Disagreement, Protocol Sensitivity, and a Matched Benchmark-Traffic Difference of at Most Three Points.
- Paper: https://openreview.net/forum?id=6ccINAFoJi
- Code and public harness: https://github.com/AmperesAI/amperes-bench
- Author: NarasingaMoorthy V
Contents
data/eval_set.jsonlcontains the 11,364-prompt evaluation set and its task and complexity labels.artifacts/tmlr_supplementary_reviewer_revision.zipcontains the scripts, saved evaluation outputs, run logs, and numerical verifier used for the camera-ready paper.SHA256SUMSrecords the checksums of both files.
Run the verifier after extracting the artifact bundle:
python3 tmlr_supp/scripts/verify_paper_numbers.py tmlr_supp/data/reports
Release boundary
The artifact excludes raw WildChat text, public-human prompt and response text, and judge reasoning. Public human labels cover external response pairs and are not labels for the paper's focal benchmark or WildChat pairs. The original WildChat source revision was not pinned, so that sample cannot be regenerated byte-for-byte.
The benchmark combines material from several public datasets. Their original terms continue to apply. The repository does not assign a new license to third-party material.
Citation
@article{v2026downshifting,
title = {How Much Quality Survives LLM Downshifting? Judge Disagreement, Protocol Sensitivity, and a Matched Benchmark--Traffic Difference of at Most Three Points},
author = {V, NarasingaMoorthy},
journal = {Transactions on Machine Learning Research},
year = {2026},
url = {https://openreview.net/forum?id=6ccINAFoJi}
}
- Downloads last month
- 33