Hugging Face's logo

Datasets:
LayerLens
/
gmail-may-invoice-count-test

Tasks:
Text Generation
Question Answering
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
agents
agent-evaluation
tool-calling
function-calling
email
gmail
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
Dataset Viewer
Auto-converted to Parquet Duplicate
run_key
string
task_id
string
agent
string
answer_line
string
answer
string
passed
bool
verdict
string
trace_status
string
tool_calls
int64
tool_calls_note
string
gmail_tool_calls
int64
key_tool_calls
list
cost_usd
float64
threads_returned_by_2025_05_search
int64
threads_opened_from_that_search
int64
thread_not_opened
string
invoice_in_thread_not_opened
string
tokens_total
int64
grading_evidence_text
string
stop_event_found
bool
may-count-a
may-2025-invoice-count
Agent A
6
null
true
pass
completed
52
count as transcribed in the 3 Nov snapshot; whether it includes tool_search loader calls is not recorded
null
null
null
null
null
null
null
null
null
null
may-count-c
may-2025-invoice-count
Agent C
6
null
true
pass
completed
43
count as transcribed in the 3 Nov snapshot; whether it includes tool_search loader calls is not recorded
null
null
null
null
null
null
null
null
null
null
may-count-e
may-2025-invoice-count
Agent E
5
5 ANSWER: 5
false
fail
completed
11
11 tool.call events: 10 Gmail calls and 1 tool_search loader call; the run panel shows 10
10
[ { "tool": "tool_search", "args": { "limit": 5, "query": "gmail search messages by invoice due date May 2025" } }, { "tool": "gmail__search_messages", "args": { "q": "invoice", "maxResults": 100 }, "resultSizeEstimate": 160 }, { "tool": "gmail__search_m...
0.008451
6
5
zzz0000000000068
INV-2025-0016, 6,550 USD, Northgate Logistics Group, due 2025-05-17
77,618
looked for "6" in the agent's final-answer line
null
may-count-b
may-2025-invoice-count
Agent B
null
false
error
error
27
count as transcribed in the 3 Nov snapshot; whether it includes tool_search loader calls is not recorded
null
null
0.41473
null
null
null
null
null
null
false
may-count-d
may-2025-invoice-count
Agent D
null
false
fail
completed
20
count as transcribed in the 3 Nov snapshot; whether it includes tool_search loader calls is not recorded
null
null
0.289073
null
null
null
null
null
null
false

Gmail May Invoice Count Test

Blog post | LayerLens

5 agents, each built on a different foundation model, ran once each on 1 question in a simulated Gmail inbox for an accounts payable team, the people who pay supplier bills: how many invoices are due in May 2025? The known answer is 6.

2 of the 5 runs answered 6 and passed. 1 answered 5 and failed, and 2 returned an empty answer. The run that answered 5 searched for "2025-05", got back 12 messages in 6 threads, opened 5 of those threads and never opened the 6th, which holds INV-2025-0016 for 6,550 USD, due 2025-05-17.

The vendors, invoices and email addresses are fictional. All 5 runs happened inside LayerLens Environments, which simulates services such as Gmail so an agent can be tested on records with known answers. This pack holds the graded run records and the tool calls of the failed count, not full traces.

Load 

from datasets import load_dataset

runs = load_dataset("LayerLens/gmail-may-invoice-count-test", "runs", split="train")
tasks = load_dataset("LayerLens/gmail-may-invoice-count-test", "tasks", split="train")

Results

Agents are labeled A to E, with the same labels in the Gmail Same-Name Vendor Test. Foundation model names are not reported, and the results are not a ranking.

Agent Final answer Grade Trace status Tool calls
Agent A 6 Pass completed 52
Agent C 6 Pass completed 43
Agent E 5 Fail completed 11 (10 Gmail)
Agent B none Error error 27
Agent D none Fail completed 20

2 of 5 passed, matching the task page in LayerLens (2 / 5 passed).

The task

How many invoices in this mailbox are due in May 2025? Answer with the number only.

Known answer: 6. The 6 invoices, as listed in the 2 passing answers:

Invoice Due
INV-2025-0013 2025-05-01
INV-2025-0014 2025-05-06
INV-2025-0015 2025-05-12
INV-2025-0016 2025-05-17
INV-2025-0017 2025-05-23
INV-2025-0018 2025-05-29

The environment

gmail-ap-inbox-hard-cases, version 1, in LayerLens Environments, task set agg count by month. An agent searches and reads a simulated Gmail mailbox through Gmail tools, which it looks up with a tool_search loader call.

The failed count, step by step

The Agent E run (run may-count-e) made 11 tool calls, 10 of them Gmail calls:

  1. tool_search for Gmail search tools.
  2. gmail__search_messages with q: "invoice": estimate of 160 results.
  3. gmail__search_messages with q: "\"May 2025\"": 0 results.
  4. gmail__search_messages with q: "invoice due": estimate of 60 results.
  5. gmail__get_thread on 1 thread holding INV-2025-0040.
  6. gmail__search_messages with q: "\"2025-05\"": 12 messages in 6 threads.
  7. to 11. gmail__get_thread on 5 of those 6 threads, holding INV-2025-0017, INV-2025-0015, INV-2025-0013, INV-2025-0018 and INV-2025-0014.

The thread it did not open holds INV-2025-0016: "Invoice INV-2025-0016 is for 6,550 USD and is due on 2025-05-17." That text comes from the same thread as read in the 2 passing runs.

Its answer:

5 ANSWER: 5

The run used 77,618 tokens and cost 0.008451 USD.

What the grader checks and what it does not

The grader is an exact-value check on the final ANSWER: line against the known answer, 6. Its logged reason: looked for "6" in the agent's final-answer line.

It does not check:

  • which invoices an agent counted. The grade depends only on the final number; the 2 passing runs happen to list the 6 invoices above.
  • why a run stopped. The trace shows which thread the failed run skipped, not why; the traces hold tool calls and results, not reasoning.
  • why 2 runs returned nothing. Neither empty trace holds a cost-cap or stop event, so the cause is not recorded.

Reproduce on LayerLens

In LayerLens Environments, open gmail-ap-inbox-hard-cases version 1, task set agg count by month, and run the task on an agent. Each attempt is graded by the exact-value check and saved as a trace with its tool calls. Whether this environment is open to accounts outside the LayerLens team is a Marin question.

Limitations

  • 5 runs, 1 per agent, on 1 question. The pack does not show how often any foundation model answers this way.
  • The 6-invoice list comes from the 2 passing answers, which agree with the known answer of 6; the pack holds no separate invoice table from the environment.
  • Tool calls are listed for the failed count only. The counts for the other 4 runs were transcribed from the records without saying whether they include tool-search loader calls.
  • Cost is in the pack for 3 runs: 0.008451 USD for the failed count, 0.41473 USD and 0.2890725 USD for the 2 empty answers.
  • These runs happened inside the LayerLens product and cannot be rerun from this pack.

Files

File Contents
runs.jsonl 1 row per run: run key, agent label, final answer, grade, trace status, tool calls, cost where recorded
tasks.jsonl The task, the known answer, the grader and the 6 invoices listed in the passing answers

Citation 

@misc{layerlens2026gmailmaycount,
  title  = {Gmail May Invoice Count Test},
  author = {{LayerLens}},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/LayerLens/gmail-may-invoice-count-test}}
}
Downloads last month
27
Number of rows:
6
Total file size:
13 kB