Dataset Viewer
Auto-converted to Parquet Duplicate
messages
listlengths
4
30
tools
listlengths
4
10
model
stringclasses
1 value
pilot
stringclasses
1 value
run
stringclasses
1 value
date
timestamp[s]date
2025-01-19 00:00:00
2025-01-19 00:00:00
desk
stringclasses
3 values
watch
int64
1
24
watches
int64
5
24
day_reward
float64
0.91
0.91
day_cost_usd
int64
1.83M
1.83M
[ { "role": "system", "content": "You are the crew scheduler of JetBlue's operations control center. The ops controller runs the flights and aircraft, with the FAA, and the customer desk looks after the passengers, two other desks; you run the crews. Keep JetBlue's cost of the day as low as you can.\n\nThe da...
[ { "type": "function", "function": { "name": "crew_flights", "description": "As ops_flights, with the crews the projection has on each flight; status attention adds every flight the crew system has a word on.", "parameters": { "additionalProperties": false, "properties": { ...
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
crew
1
5
0.908512
1,829,878
[ { "role": "system", "content": "You are the crew scheduler of JetBlue's operations control center. The ops controller runs the flights and aircraft, with the FAA, and the customer desk looks after the passengers, two other desks; you run the crews. Keep JetBlue's cost of the day as low as you can.\n\nThe da...
[ { "type": "function", "function": { "name": "crew_flights", "description": "As ops_flights, with the crews the projection has on each flight; status attention adds every flight the crew system has a word on.", "parameters": { "additionalProperties": false, "properties": { ...
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
crew
2
5
0.908512
1,829,878
[ { "role": "system", "content": "You are the crew scheduler of JetBlue's operations control center. The ops controller runs the flights and aircraft, with the FAA, and the customer desk looks after the passengers, two other desks; you run the crews. Keep JetBlue's cost of the day as low as you can.\n\nThe da...
[ { "type": "function", "function": { "name": "crew_flights", "description": "As ops_flights, with the crews the projection has on each flight; status attention adds every flight the crew system has a word on.", "parameters": { "additionalProperties": false, "properties": { ...
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
crew
3
5
0.908512
1,829,878
[ { "role": "system", "content": "You are the crew scheduler of JetBlue's operations control center. The ops controller runs the flights and aircraft, with the FAA, and the customer desk looks after the passengers, two other desks; you run the crews. Keep JetBlue's cost of the day as low as you can.\n\nThe da...
[ { "type": "function", "function": { "name": "crew_flights", "description": "As ops_flights, with the crews the projection has on each flight; status attention adds every flight the crew system has a word on.", "parameters": { "additionalProperties": false, "properties": { ...
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
crew
4
5
0.908512
1,829,878
[ { "role": "system", "content": "You are the crew scheduler of JetBlue's operations control center. The ops controller runs the flights and aircraft, with the FAA, and the customer desk looks after the passengers, two other desks; you run the crews. Keep JetBlue's cost of the day as low as you can.\n\nThe da...
[ { "type": "function", "function": { "name": "crew_flights", "description": "As ops_flights, with the crews the projection has on each flight; status attention adds every flight the crew system has a word on.", "parameters": { "additionalProperties": false, "properties": { ...
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
crew
5
5
0.908512
1,829,878
[ { "role": "system", "content": "You are the customer desk of JetBlue's operations control center. The ops controller runs the flights and aircraft, with the FAA, and the crew scheduler runs the crews, two other desks; you look after the passengers. Keep JetBlue's cost of the day as low as you can.\n\nThe da...
[ { "type": "function", "function": { "name": "cust_flights", "description": "Flights by scheduled departure, as many as fit 4k characters, filtered by any of ids, station, status and scheduled departure window: booked passengers and of them connecting onto and off it, where the projection has it ...
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
customer
1
24
0.908512
1,829,878
[{"role":"system","content":"You are the customer desk of JetBlue's operations control center. The o(...TRUNCATED)
[{"type":"function","function":{"name":"cust_flights","description":"Flights by scheduled departure,(...TRUNCATED)
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
customer
2
24
0.908512
1,829,878
[{"role":"system","content":"You are the customer desk of JetBlue's operations control center. The o(...TRUNCATED)
[{"type":"function","function":{"name":"cust_flights","description":"Flights by scheduled departure,(...TRUNCATED)
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
customer
3
24
0.908512
1,829,878
[{"role":"system","content":"You are the customer desk of JetBlue's operations control center. The o(...TRUNCATED)
[{"type":"function","function":{"name":"cust_flights","description":"Flights by scheduled departure,(...TRUNCATED)
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
customer
4
24
0.908512
1,829,878
[{"role":"system","content":"You are the customer desk of JetBlue's operations control center. The o(...TRUNCATED)
[{"type":"function","function":{"name":"cust_flights","description":"Flights by scheduled departure,(...TRUNCATED)
reflection/Beam-1.0-501B-A23B:baseten
occ3-real-beam
b6-20250119-0400-joint
2025-01-19T00:00:00
customer
5
24
0.908512
1,829,878
End of preview. Expand in Data Studio

AirOpsEnv x Beam

Reflection's Beam runs JetBlue's operations control center through a real storm day, January 19, 2025. It plays all three desks: flight operations, crew scheduling and customer service. Each desk is a separate tool-calling agent that can only use its own role's tools and has to ask the other desks for anything else. We score the plan on what it would have cost JetBlue that day, and we turn every desk watch Beam played into a reasoning-annotated row for supervised fine-tuning.

This repository holds:

  • sft/train.jsonl: 53 training rows, one per desk watch. Each row has the desk's rules, that hour's page, and Beam's tool calls with the reasoning_content it wrote before each one. The rows are in the OpenAI chat format with a tools column, which TRL's SFTTrainer reads directly.
  • run/: the run itself. That's the result row with the reward and the audit, the scored day record with the plan and its cost broken down, and each desk's transcript.

The task

When thunderstorms or snow hit New York and Boston, an airline's control center decides which departures to hold, which flights to cancel, where to move crews and how to protect connecting passengers. AirOpsEnv is an AgentEnv environment that rebuilds JetBlue's day from public records:

  • Flights: the Bureau of Transportation Statistics on-time data.
  • FAA programs: ground stops and ground delay programs from the FAA Command Center, released to the desks when they were issued.
  • Weather: METAR and TAF reports and NEXRAD radar from the Iowa Environmental Mesonet.

Crews and passengers are generated stand-ins. Pilots follow FAR 117 duty limits, and passenger itineraries come from the DOT's T-100 and Origin and Destination data, about 86,000 trips on a storm date. Stranded passengers are rebooked under JetBlue's Contract of Carriage.

This run uses the environment's realistic mode, which adds the limits an airline actually runs into:

  • hourly airport capacity built from every carrier's recorded pushbacks and landings;
  • a parking limit at each station;
  • aircraft that break down at JetBlue's fair-weather rate;
  • aircraft swaps only at JFK and Boston;
  • a fleet that has to end the day in position for tomorrow's first departures.

The desks work in hourly watches from 04:00. A flight's delay becomes visible two hours ahead, and no decision can reach back in time.

Desk Can do Asks the others
Ops controller cancel, hold, claim and swap FAA arrival slots COVER this flight
Crew scheduler call reserves, swap and deadhead crews, book hotels HOLD this flight, CANCEL this round trip
Customer desk look up connections and who would be stranded HOLD for connections, PROTECT a flight, SUGGEST_CANCEL one

Scoring

The objective is to minimize JetBlue's financial loss for the day. The cost counts:

  • aircraft delay at the FAA's cost rates, less the fuel and maintenance a cancellation saves;
  • crew deadheads and hotels;
  • tomorrow's late starts;
  • the fares lost on refunded and stranded passengers.

The two free parameters are fitted to JetBlue's Q1 2014 filing. A replay of JetBlue's own decisions scores 0.5, and the cheapest plan found with hindsight scores 1.0. A run counts only if an audit confirms that each desk stayed in its role, saw nothing early and left every flight with a legal crew.

How Beam was used

reflection/Beam-1.0-501B-A23B plays every desk, called through Hugging Face Inference Providers (reflection/Beam-1.0-501B-A23B:baseten, the OpenAI-compatible router). Every desk ran at reasoning effort high.

  • Tool calling. Each desk gets only its own role's tools, and the environment turns off everything else. Beam reads the page for the hour, calls tools to look things up and act, sends typed requests to the other desks, and ends its watch with end_turn.
  • Sending reasoning back. Beam rejects a conversation unless every earlier assistant turn carries the reasoning_content Beam wrote for it. AirOpsEnv's desk agent now sends each turn's reasoning back on every request, using a short placeholder for the rare turn that came without any. Beam's thinking therefore carries across the whole watch. We added this for the entry (commit e6e573b), with tests.
  • Cost caps. Each desk was capped at $7 of model spend, plus $1 for one last request. A desk that reaches its cap stops acting, and the day still scores. Beam spent $3.74 over 12 minutes.
  • Beam as teacher. scripts/sft_rows.py turns each desk watch into a training row. Rows keep Beam's reasoning on every assistant turn and the tools that role actually had. The messages leave out the next hour's page and the day's reward, so a model trained on them never sees what the desk couldn't have known; the reward sits in its own column for filtering. A watch with a turn cut off at the output limit, or a turn without reasoning, is left out.

Beam on January 19, 2025

Plan Reward Cost to JetBlue Flights not flown
Beam, three desks 0.909 $1.83M 19
GPT-6 Astra, three desks 0.940 $1.71M 21
Scripted standard desks $1.88M
JetBlue's own decisions, replayed 0.5 $3.28M 157 cancelled (recorded)
Cheapest plan found with hindsight 1.0 $1.27M 35

Beam's three desks cost JetBlue $1.83M on the day: 3% less than the scripted standard desks, 44% less than JetBlue's replayed decisions and 7% more than GPT-6 Astra's three desks. Beam left 19 flights unflown, against 21 for Astra and 35 for the hindsight plan. The ops and customer desks played all 24 watches; the crew desk stood down after its fifth, which is why it has 5 of the 53 rows.

On this date JetBlue's replayed decisions cost at least two-thirds more than any other plan we priced, so every plan that keeps flying scores well. The useful comparisons are with the scripted standard desks and with the hindsight plan. One day is one sample. Treat these numbers as a demonstration, not a benchmark.

Train with TRL

from datasets import load_dataset
from trl import SFTConfig, SFTTrainer

ds = load_dataset("agutmanstein-scale/AirOpsEnv-Beam", split="train")

trainer = SFTTrainer(
    model="Qwen/Qwen3-8B",
    train_dataset=ds,
    args=SFTConfig(output_dir="airops-desk-sft", max_length=32768, per_device_train_batch_size=1,
                   gradient_accumulation_steps=8, num_train_epochs=2, bf16=True),
)
trainer.train()

Each row has messages, with system, user, assistant and tool turns, and with reasoning_content on assistant turns. It also has tools, the role's function schemas. Use a chat template that renders reasoning_content, as Qwen3's does, so the model learns the thinking along with the calls. The metadata columns are desk, watch, watches, date, day_reward and day_cost_usd. You can use them to filter by desk, or to weight rows by how well the day went.

Reproduce

git clone https://github.com/agutmanstein-scale/agentenv-airops-plugin -b beam/submission && cd agentenv-airops-plugin
uv sync --all-extras && uv pip install agentenv-framework==0.9.1339 agentenv-framework-protocol==0.1.354
.venv/bin/agent-env airops-occ3 setup
.venv/bin/agent-env airops-occ3 tasks --realistic --split all
export LITELLM_BASE_URL=https://router.huggingface.co LITELLM_API_KEY=<your HF token with Inference Providers access>
.venv/bin/agent-env run results/bundles/airops-occ3-real --task b6-20250119-0400-joint-real \
    --model reflection/Beam-1.0-501B-A23B:baseten
.venv/bin/python scripts/sft_rows.py results/runs/<pilot> sft/train.jsonl

Credits and license

AirOpsEnv was built by Edgar Arakelyan, and realistic mode and this Beam entry by Alec Gutmanstein, both at Scale AI. The environment code is Apache-2.0, and Beam is Apache-2.0. The flight, FAA and weather inputs are public US government and Iowa State University records. Crews and passengers are synthetic. This dataset is released under Apache-2.0.

Downloads last month
105