kuririrn commited on
Commit
38da88e
·
verified ·
1 Parent(s): a71cad9

Add dataset card for mixed ALFWorld trajectories

Browse files
Files changed (1) hide show
  1. README.md +42 -21
README.md CHANGED
@@ -1,23 +1,44 @@
1
  ---
2
- dataset_info:
3
- features:
4
- - name: messages
5
- list:
6
- - name: role
7
- dtype: string
8
- - name: content
9
- dtype: string
10
- - name: dataset_version
11
- dtype: string
12
- splits:
13
- - name: train
14
- num_bytes: 11039064.0
15
- num_examples: 2359
16
- download_size: 2138461
17
- dataset_size: 11039064.0
18
- configs:
19
- - config_name: default
20
- data_files:
21
- - split: train
22
- path: data/train-*
23
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ datasets:
3
+ - kuririrn/sft_alfworld_trajectory_dataset_v3to5_admissible_plus_v5extra500
4
+ pretty_name: ALFWorld SFT Trajectories (v3–v5 Admissible + v5 Extra 500)
5
+ tags:
6
+ - alfworld
7
+ - sft
8
+ - text-generation
9
+ - tool-use
10
+ - reinforcement-learning
11
+ language:
12
+ - en
13
+ license: other
14
+ task_categories:
15
+ - text-generation
16
+ task_ids:
17
+ - conversational
18
+ size_categories:
19
+ - 1K<n<10K
 
 
 
20
  ---
21
+
22
+ # Dataset Card for ALFWorld SFT Trajectories (v3–v5 Admissible + v5 Extra 500)
23
+
24
+ ## Overview
25
+
26
+ This dataset extends the **Admissible-only** subset by adding a fixed
27
+ number of trajectories without admissible actions from v5.
28
+
29
+ It is constructed as follows:
30
+
31
+ 1. Take all trajectories (episodes) from:
32
+ - `u-10bei/sft_alfworld_trajectory_dataset_v3`
33
+ - `u-10bei/sft_alfworld_trajectory_dataset_v4`
34
+ - `u-10bei/sft_alfworld_trajectory_dataset_v5`
35
+ that contain "admissible actions" in at least one observation
36
+ message. (Same as the admissible-only dataset.)
37
+ 2. Additionally, sample 500 trajectories from
38
+ `u-10bei/sft_alfworld_trajectory_dataset_v5` that do **not** contain
39
+ admissible actions.
40
+ 3. Concatenate them and shuffle.
41
+
42
+ This dataset can be used to evaluate the robustness of an SFT model
43
+ trained mainly on admissible-constrained trajectories, when exposed to
44
+ a mix of constrained and unconstrained episodes.