amitnaik07/beam-data-desk-traces

Beam Data Desk traces

The tasks and full agent traces behind Beam Data Desk, a benchmark of tool-using agents answering business questions over a synthetic B2B SaaS database. Reflection Beam 1.0, DeepSeek-V4.1-Flash and GLM-5.3 each answered all 67 questions with five read-only tools. Every run is here with its answer, its score and the complete step-by-step trace: the model's reasoning, each tool call and each tool result.

Use it to study how agents fail on realistic analytics questions, to compare reasoning-effort settings, or to score your own agent on the same tasks.

What's in it

Config	Rows	One row is
runs (default)	469	one model answering one task at one reasoning effort
tasks	67	one question with its gold answer, the SQL that computes it, and notes on its trap
experiments	132	a follow-up run: a repeat of Beam on a hard task, or Beam high with a metric definition (or a placebo) added to the prompt

data/desk.db is the SQLite database the questions are asked against (120 accounts, 344 opportunities, 656 tickets, 910 invoices and more, all synthetic). Every question is answered as of 2026-10-01.

The 469 runs cover these settings, 67 tasks each:

Model	Efforts
Reflection Beam 1.0 (reflection/Beam-1.0-501B-A23B, on Baseten)	low, medium, high
DeepSeek-V4.1-Flash-1017 (on Baseten)	low, max
GLM-5.3 (zai-org/GLM-5.3, on DeepInfra)	low, high

from datasets import load_dataset

runs = load_dataset("amitnaik07/beam-data-desk-traces", "runs", split="train")
tasks = load_dataset("amitnaik07/beam-data-desk-traces", "tasks", split="train")
exp = load_dataset("amitnaik07/beam-data-desk-traces", "experiments", split="train")

misses = runs.filter(lambda r: r["model_label"] == "Beam 1.0" and not r["correct"])

Results

Model	Effort	Correct (of 67)	Hard tasks (of 20)	Avg tool calls	Median reply	Cost per correct answer
Beam 1.0	low	53	6	4.0	1.8 s	$0.0074
Beam 1.0	medium	60	13	4.9	2.4 s	$0.0094
Beam 1.0	high	62	15	5.1	2.4 s	$0.0109
DeepSeek-V4.1-Flash	low	63	16	4.9	1.9 s	$0.0083
DeepSeek-V4.1-Flash	max	64	17	7.1	2.0 s	$0.0164
GLM-5.3	low	47	2	3.1	5.0 s	$0.0127
GLM-5.3	high	51	4	3.6	4.1 s	$0.0153

All three models get the 47 core tasks right (GLM low misses 2). The 20 hard tasks separate them. Beam improves at every effort step, and all five of Beam high's misses come from one definition: it computes gross revenue retention as a plain revenue ratio (or as net retention) instead of capping each customer at last year's revenue. No run gave a value that appears nowhere in its tool output or reasoning. The full analysis is in the Space's REPORT.md.

Follow-up experiments

The experiments config holds two follow-ups run on 2026-10-31, kept out of runs so that config stays one run per task and setting.

• Repeats (experiment = "repeat"): Beam medium and high ran the 20 hard tasks twice more. With the first sweep that makes three runs each: medium scored 13, 18 and 14, high 15, 17 and 15.
• Definition (experiment = "definition"): Beam high ran the 13 tasks that depend on gross revenue retention with one paragraph added to the system prompt, two runs each. grr_def states the per-account GRR rule; placebo defines average days to pay, which no task uses.

System prompt	Correct of 13, per run	Mean
normal (first sweep plus the two repeats)	8, 11, 9	9.3
+ GRR definition	12, 13	12.5
+ placebo	8, 10	9.0

Paired run by run, the definition was right where the placebo was wrong 7 times and never the reverse (exact McNemar p = 0.016), with 31% fewer output tokens. The paragraph texts are in desk/scoring.py (PROMPT_ADDENDA) in the Space repo.

Fields

runs

Field	Meaning
run_id, task_id	ids; join to tasks on task_id
task_set	core (T01-T47), hard (T48-T57) or harder (T58-T67)
category	lookup, aggregation, join, multi_hop, disambiguation, temporal, safety, hard or harder
model, model_label, effort	router model id with provider, display name, reasoning_effort sent
correct	whether the scorer accepted the answer
answer, gold	the parsed FINAL ANSWER and the gold answer, as strings (lists are JSON)
score_reason	the scorer's explanation, such as missing ['audit log']
error	set when a run ended early, such as hitting the step limit
tool_calls, prompt_tokens, completion_tokens, reasoning_tokens, latency_s	usage for the whole run
cost_usd	list-price estimate from the token counts (prices checked 2026-10-22)
final_text	the model's closing message
steps_json	the trace as a JSON list of reasoning, tool_call, tool_result and final events; tool results are cut at 1,500 characters

experiments has every runs field plus experiment (repeat or definition), prompt (base, grr_def or placebo) and run_no.

tasks

Field	Meaning
task_id, task_set, category, question	the question as the agent saw it
answer_type	number, string, set or refusal
gold, tolerance	the gold answer (as a string) and the relative tolerance used for numbers
gold_sql	the SQL that computes the gold answer from desk.db
trap_notes	what makes the task hard and which wrong answer each trap produces

How runs were made and scored

Each run gave the model the question, a system prompt and five tools: list_tables, describe_table, run_sql (one read-only SELECT), search_tickets and get_account. The agent loop ran for at most 24 steps, and the model had to end with a line FINAL ANSWER: <value>. (The first sweeps used 12; the 13 runs that hit that limit were rerun with 24 and replace the originals here. Every other run finished within 12 steps.) Numbers are compared within a tolerance, sets as sets, and requests to change data or dump personal data must be answered REFUSED. Runs went through the Hugging Face router in October 2026, one run per task and setting in runs, so a difference of one or two tasks between settings is within run-to-run noise (the repeats in experiments show how much).

The code that builds the database, the tasks and these files is in the Space repo (desk/build_db.py, desk/build_tasks.py, eval/export_dataset.py).

Limits

• The hard tasks were picked by screening candidates on these three models and keeping the ones that broke at least one. They show where these models differ, not how often a typical question goes wrong.
• 13 of the 20 hard tasks depend on gross revenue retention.
• Effort levels are each provider's own scale, not equal token budgets.
• Reply times include provider queueing.

All companies, people and numbers are synthetic. Any resemblance to real ones is a coincidence.