flickery/pasarbench-traces

PasarBench traces

Every recorded episode of PasarBench, a verifiable customer-service agent benchmark for Southeast Asian e-commerce: 22,191 episodes in 57 runs, 69 run cells, 8 language varieties (en, id, ms, sg-en, th, vi, zh-MY, zh-SG); and 2,691 continuations of the language swap in 5 runs, below. The code, the tasks and the write-up of every run are in https://github.com/FlickYan/pasarbench; each run's question and result is a section of its docs/RUNBOOK.md.

An episode is one conversation between an agent model and a simulated customer, with tools over a mock order database. It passes when the database ends in the right state -- the refund issued, the escalation filed, the forbidden action never taken -- not when the reply reads well.

Agent models

agent model	episodes
deepseek-v4-pro	14,150
Qwen/Qwen3.8-27B	2,936
Qwen/Qwen3.8-27B:pasar-grpo2-i1-B	790
Qwen/Qwen3.8-27B:pasar-rft2-B	790
Qwen/Qwen3.8-27B:pasar-grpo2-i1-A	775
Qwen/Qwen3.8-27B:pasar-rft2-A	775
Qwen/Qwen3.8-27B:pasar-rft-B	550
Qwen/Qwen3.8-27B:pasar-rft-A	525
deepseek-flash	225
glm-5.3-flash	225
grok-4.3	225
mimo-v2.6-flash	225

Fields

field	what it is
run, cell	the run (one experiment) and its cell (one setup within it)
episode	the trace's file name: task id and repetition
task_id, trap, market, language	the task, the policy trap it tests, its market and the language the customer writes in
agent_model, simulator	the agent under test and the simulated customer
policy_mode, exposure, context	how the policy and the tools were shown, and how the conversation was kept
closing, guardrail	the closing check and the claim guardrail, when a run had them
passed, failures	the verdict the run recorded, and the checks it failed
stop_reason, steps, tokens	how the episode ended and what it used
trace	the whole trace, one JSON record a line: header, steps (each model call, its tool calls and whether each succeeded), events (customer turns, notes) and footer

Read this before using the verdicts

• passed is the verdict of the checker the day the run happened. Checks corrected since (docs/WHAT_FAILED.md) change some of them; scripts/rescore.py replays an episode's calls and scores it under today's checks.
• Runs differ in setup on purpose: the tool exposure, the policy mode, the closing check, and the simulated customer's rules (its version 2 from v31 on). Compare cells only within a run, paired by task, as the runbook does.
• The customers and every customer record are synthetic. Email addresses written in the conversations are replaced with [email] (5 in 4 episodes).

The language swap (RUNBOOK 1t, 1v, 1w)

2,691 continuations in 5 runs. Each holds a recorded duplicate-refund conversation fixed at the moment the agent reads a refund already paid -- the system prompt, every message, every tool payload and the world as they were -- and lets the agent continue, in arms that change only the conversation's language (1t; 1w, on other agents), or the policy's language too (1v), until it next speaks to the customer. The outcome is escalated: whether it escalated before that reply, which the policy's P10 asks of a refund already paid. There is no passed.

run	RUNBOOK	agent model	arm	continuations	escalated	what it was shown, checked
V-policy	1v	deepseek-v4-pro	english	288	130 (45%)	rebuilt; no digest before v70
V-policy	1v	deepseek-v4-pro	local	288	62 (22%)	rebuilt; no digest before v70
V-policy	1v	deepseek-v4-pro	policy	288	127 (44%)	rebuilt; no digest before v70
W-swap	1t	deepseek-v4-pro	back	288	95 (33%)	rebuilt; no digest before v70
W-swap	1t	deepseek-v4-pro	orig	288	88 (31%)	rebuilt; no digest before v70
W-swap	1t	deepseek-v4-pro	swap	288	94 (33%)	rebuilt; no digest before v70
X-deepseek-flash	1w	deepseek-flash	back	108	97 (90%)	rebuilt; no digest before v70
X-deepseek-flash	1w	deepseek-flash	orig	108	93 (86%)	rebuilt; no digest before v70
X-deepseek-flash	1w	deepseek-flash	swap	108	99 (92%)	rebuilt; no digest before v70
X-glm-5.3-flash	1w	glm-5.3-flash	back	105	76 (72%)	rebuilt; no digest before v70
X-glm-5.3-flash	1w	glm-5.3-flash	orig	105	76 (72%)	rebuilt; no digest before v70
X-glm-5.3-flash	1w	glm-5.3-flash	swap	105	82 (78%)	rebuilt; no digest before v70
X-mimo-v2.6-flash	1w	mimo-v2.6-flash	back	108	102 (94%)	rebuilt; no digest before v70
X-mimo-v2.6-flash	1w	mimo-v2.6-flash	orig	108	104 (96%)	rebuilt; no digest before v70
X-mimo-v2.6-flash	1w	mimo-v2.6-flash	swap	108	100 (93%)	rebuilt; no digest before v70

Config swap, one line per continuation:

field	what it is
run, experiment	the run, and the RUNBOOK section it answers
arm	orig, swap, back (1t, 1w); local, policy, english (1v); each of those with -plain or -named (1x)
pid	the prefix: the recorded conversation held fixed, the same in every arm and run that continued it
task_id, run_index, market, trap	the task, the continuation's repetition, its market and trap
agent_model, translator	the agent continuing, and the model that translated the conversation and the policy
recorded_language, target, language, local_language, policy_language	the language it was recorded in and swapped into; this arm's conversation; the language other than English; the policy's
tools_named, exposure	whether this arm's policy named the tools; the tools the run showed: every one (default) or behind search
source_run, source_cell, source_episode	the recorded conversation, in the episodes config
prefix_steps, recorded_escalated	the agent's steps up to the lookup; whether the recording escalated before its next reply
escalated, category, calls, reply, searched_nothing	escalated before its next reply, the category it gave, the tools it called, what it then said, policy searches that found nothing
stop_reason, steps, tokens, started_at	how it ended and what it used
trace	the continuation's whole trace: its header (the arm under swap), steps and footer

Config swap_prompts, one line per arm of each prefix in a run -- join on run, pid and arm: run, experiment, arm, pid, task_id, language, policy_language, tools_named and translator as above, and messages, JSON: every message the agent was shown, in order -- the system prompt with its policy, her turns and its replies in the arm's language, every tool call and payload.

• The translations are a model's, qwen3.8-flash at temperature 0 behind one cache: checked for the codes, numbers, sections and emphasis each must keep, and read before the runs; not reviewed by native speakers.
• messages is rebuilt from the prefixes and the run's own translation cache, as the run built it, with no call made. Each continuation from v70 records a digest of what it was shown, and the export checks it; for the earlier runs a translated policy is checked against the text the run wrote beside its traces.
• Pair the arms by pid within a run. The reading fixed before each run, and its result, are in docs/RUNBOOK.md (1t, 1v, 1w, 1x) and RESULTS.md section 9. Email addresses are replaced here too (60).

Load

from datasets import load_dataset
eps = load_dataset("<this dataset>", "episodes", split="train")
import json; records = [json.loads(l) for l in eps[0]["trace"].splitlines()]
swap = load_dataset("<this dataset>", "swap", split="train")
shown = load_dataset("<this dataset>", "swap_prompts", split="train")
messages = json.loads(shown[0]["messages"])

License: MIT, as the code.