Hugging Face's logo

Datasets:
violetxi
/
qwen35-4b-tmax-16turn-unfinished-trajectories

Modalities:
Tabular
Text
Formats:
parquet
Size:
< 1K
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
messages
listlengths
32
32
task_name
stringclasses
11 values
attempt
int32
1
16
turn_count
int32
16
16
binary_reward
float64
0
1
verifier_completed
bool
1 class
termination_reason
stringclasses
1 value
source_job_id
stringclasses
1 value
model
stringclasses
1 value
temperature
float64
0.6
0.6
top_p
float64
1
1
max_turns
int32
16
16
max_tokens_per_turn
int32
8.19k
8.19k
[ { "role": "user", "content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON w...
task_000877_cc92f09d
1
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[ { "role": "user", "content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON w...
task_000877_cc92f09d
2
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[ { "role": "user", "content": "You are an AI assistant tasked with solving command-line tasks in a Linux environment. You will be given a task description and the output from previously executed commands. Your goal is to solve the task by providing batches of shell commands.\n\nFormat your response as JSON w...
task_000877_cc92f09d
3
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED)
task_000877_cc92f09d
4
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED)
task_000877_cc92f09d
5
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED)
task_000877_cc92f09d
6
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED)
task_000877_cc92f09d
7
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED)
task_000877_cc92f09d
8
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED)
task_000877_cc92f09d
9
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
[{"role":"user","content":"You are an AI assistant tasked with solving command-line tasks in a Linux(...TRUNCATED)
task_000877_cc92f09d
10
16
0
true
max_turns_without_task_complete
891542
Qwen/Qwen3.5-4B
0.6
1
16
8,192
End of preview. Expand in Data Studio

Qwen3.5-4B TMAX 16-turn unfinished trajectories

This public, ungated dataset contains the 77 trajectories from TACC Slurm job 891542 that exhausted all 16 allowed turns without the model emitting a task_complete=true action. They were selected from 256 trajectories over 16 TMAX tasks.

Message format

messages is the canonical conversational trajectory. Every row contains exactly 32 messages, strictly alternating between objects of the following form:

{"role": "user", "content": "..."}
{"role": "assistant", "content": "..."}

Only the roles user and assistant occur. Assistant messages retain the model's captured thinking and action content; user messages retain the task prompt and successive terminal observations.

Selection and caveat

  • Model: Qwen/Qwen3.5-4B
  • Sampling: temperature 0.6, top-p 1.0
  • Limits: 16 turns and 8,192 completion tokens per turn
  • Selection: turn_count == 16 and no turn has task_complete=true
  • Rows with verifier reward 1: 5/77

The absence of task_complete=true is a model-declared completion condition, not the verifier label. Five selected trajectories received binary verifier reward 1 despite reaching the turn horizon without declaring completion.
Downloads last month
37
Number of rows:
77
Total file size:
631 kB