Datasets:
messages listlengths 7 44 | system_prompt stringclasses 1
value | uid int64 0 31 | generation_id int64 0 7 |
|---|---|---|---|
[
{
"role": "user",
"content": "Hi, I’d like to make some changes to my upcoming flight. Can you help me with that?"
},
{
"role": "assistant",
"content": "The user is asking about flight changes, but no specific details were provided. I need to collect necessary information before proceeding—user ... | You are a helpful customer service agent.
<policy>
# Airline Agent Policy
The current time is 2024-05-15 15:00:00 EST.
As an airline agent, you can help users **book**, **modify**, or **cancel** flight reservations. You also handle **refunds and compensation**.
Before taking any actions that update the booking data... | 11 | 0 |
[
{
"role": "user",
"content": "Hi, I’d like to make a change to my upcoming return flight. I want to switch from my current one stop flight from Atlanta to Los Angeles, to a nonstop flight from Atlanta to Las Vegas instead. Can you help me with that?"
},
{
"role": "assistant",
"content": "The use... | You are a helpful customer service agent.
<policy>
# Airline Agent Policy
The current time is 2024-05-15 15:00:00 EST.
As an airline agent, you can help users **book**, **modify**, or **cancel** flight reservations. You also handle **refunds and compensation**.
Before taking any actions that update the booking data... | 18 | 1 |
[
{
"role": "user",
"content": "Hi! I’d like to make some changes to my upcoming flights. Can you help me with that?"
},
{
"role": "assistant",
"content": "The user wants to change their flights, but no specific details are provided. I need to ask for the user ID, reservation ID, and what changes ... | You are a helpful customer service agent.
<policy>
# Airline Agent Policy
The current time is 2024-05-15 15:00:00 EST.
As an airline agent, you can help users **book**, **modify**, or **cancel** flight reservations. You also handle **refunds and compensation**.
Before taking any actions that update the booking data... | 26 | 5 |
[
{
"role": "user",
"content": "Hi, I was told by a customer support representative over the phone that a service agent could help me cancel my reservation. I’d like to go ahead with that."
},
{
"role": "assistant",
"content": "The user wants to cancel a reservation. I need to gather specific deta... | You are a helpful customer service agent.
<policy>
# Airline Agent Policy
The current time is 2024-05-15 15:00:00 EST.
As an airline agent, you can help users **book**, **modify**, or **cancel** flight reservations. You also handle **refunds and compensation**.
Before taking any actions that update the booking data... | 24 | 0 |
[
{
"role": "user",
"content": "Hi! I have a flight booked from Atlanta to Philadelphia, but I’d like to change it to the day after my original reservation. I’m looking for the cheapest option in Economy class, but not Basic Economy. Can you help me with that?"
},
{
"role": "assistant",
"content":... | You are a helpful customer service agent.
<policy>
# Airline Agent Policy
The current time is 2024-05-15 15:00:00 EST.
As an airline agent, you can help users **book**, **modify**, or **cancel** flight reservations. You also handle **refunds and compensation**.
Before taking any actions that update the booking data... | 13 | 0 |
[
{
"role": "user",
"content": "Hi! I’d like to make some changes to my upcoming reservation. I want to move my outgoing flight to a nonstop flight on the next day, and also shift my return from SFO back by one day. Can you help me with that?"
},
{
"role": "assistant",
"content": "The user is aski... | You are a helpful customer service agent.
<policy>
# Airline Agent Policy
The current time is 2024-05-15 15:00:00 EST.
As an airline agent, you can help users **book**, **modify**, or **cancel** flight reservations. You also handle **refunds and compensation**.
Before taking any actions that update the booking data... | 0 | 7 |
[
{
"role": "user",
"content": "Hi, I’d like to book a one-way flight from New York to Seattle on May 20th."
},
{
"role": "assistant",
"content": "The user wants to book a flight. I need to collect detailed user information to proceed with the booking. First, I need the user's ID. Then, I must con... | You are a helpful customer service agent.
<policy>
# Airline Agent Policy
The current time is 2024-05-15 15:00:00 EST.
As an airline agent, you can help users **book**, **modify**, or **cancel** flight reservations. You also handle **refunds and compensation**.
Before taking any actions that update the booking data... | 15 | 3 |
[
{
"role": "user",
"content": "Hi, I need to cancel a flight that I booked about 10 hours ago. I made the reservation by mistake and I’d like to get a full refund, please."
},
{
"role": "assistant",
"content": "The user wants to cancel a flight booking made about 10 hours ago. I need to request s... | You are a helpful customer service agent.
<policy>
# Airline Agent Policy
The current time is 2024-05-15 15:00:00 EST.
As an airline agent, you can help users **book**, **modify**, or **cancel** flight reservations. You also handle **refunds and compensation**.
Before taking any actions that update the booking data... | 8 | 1 |
[{"role":"user","content":"Hi, I need help changing my return flight for my upcoming Houston to Denv(...TRUNCATED) | "You are a helpful customer service agent.\n\n<policy>\n# Airline Agent Policy\n\nThe current time i(...TRUNCATED) | 31 | 1 |
[{"role":"user","content":"Hi, I have an upcoming trip from New York to Chicago, and I’d like to m(...TRUNCATED) | "You are a helpful customer service agent.\n\n<policy>\n# Airline Agent Policy\n\nThe current time i(...TRUNCATED) | 21 | 0 |
End of preview. Expand in Data Studio
Act-PRM Thought Generations — tau2-bench Airline
Latent thoughts inferred behind logged, action-only agent demonstrations by the
Act-PRM offline EM (Action Process Reward Models), for the tau2-bench airline
domain. For each logged (state s, action x) the EM-trained generator samples
G=4 candidate thoughts z, and each candidate is scored by the
length-penalized action likelihood of the logged action:
likelihood = p(x | s, z) # policy per-action-token likelihood
reward = likelihood - 0.15 * (thought_len / 96) # reward_method="penalty"
best = argmax(reward) # top-1 selection
Generator = Qwen3-4B-Instruct-2507 LoRA. Generation params: group_size=4,
length_penalty=0.15, max_thought_tokens=96, thought_temperature=1.0,
reward_method="penalty".
Contents
generations/ — the full candidate pool (one JSONL row per logged action-step).
This is the primary artifact: it contains all G thoughts with their scores, so you
can select top-1 or recompute any weighting (em / top_half / grpo / uniform) yourself.
| file | scorer | EM checkpoint |
|---|---|---|
generations/policy.jsonl |
policy p(x|s,z) | EM step_best |
generations/base.jsonl |
base model (LoRA-detached) p(x|s,z) | EM step_best |
generations/policy_last.jsonl |
policy | EM step_last |
generations/base_last.jsonl |
base model | EM step_last |
Row schema:
{
"split": "train"|"eval", "sample_id": int, "timestep": int, "batch_id": int, "try_step": int,
"advantage_mode": "best",
"target_action": str, # the logged (ground-truth) action being explained
"thoughts": [str × G], # the G candidate thoughts
"likelihoods": [float × G], # p(x | s, z) per candidate
"rewards": [float × G], # likelihood - 0.15*len_frac per candidate
"advantages": [float × G], # per advantage_mode (here "best" -> one-hot on best)
"thought_tokens":[int × G], # candidate thought length in tokens
"best": int # argmax(rewards) — the top-1 index
}
Selecting thoughts
- Top-1 (what the paper's SFT uses):
thoughts[best]. - EM posterior: softmax-style group-normalized weights over
rewards/likelihoods. - top_half / grpo / uniform: derive from
rewards(grpo:reward - mean, opt. ÷ std).
sft_corpus_top1/ — convenience: the baked top-1 Stage-2 SFT corpus in
chat-messages format (assistant content = thoughts[best] + logged action), split
into train.json / eval.json per variant. Equivalent to taking thoughts[best] from
generations/ and joining onto the logged trajectories.
Splits
21 train trajectories / 4 eval trajectories (the rl_eval hold-out is carved out
separately for the RL phase). 258 action-step rows per variant.
Related
- SFT LoRA adapters:
mzio/aprm-sft-tau2-airline - Source expert demos:
mzio/aprm-tau2-airline-gpt5m_med-gs4-s0-train
- Downloads last month
- 23