msrewa commited on
Commit
bd58497
·
verified ·
1 Parent(s): 30f8974

Model card: finetune relation, no "expert" wording; demo video for the preview widget

Browse files
Files changed (3) hide show
  1. .gitattributes +1 -0
  2. README.md +4 -3
  3. replay.mp4 +3 -0
.gitattributes CHANGED
@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  assets/architecture.png filter=lfs diff=lfs merge=lfs -text
37
  assets/hero_teaser.webp filter=lfs diff=lfs merge=lfs -text
 
 
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  assets/architecture.png filter=lfs diff=lfs merge=lfs -text
37
  assets/hero_teaser.webp filter=lfs diff=lfs merge=lfs -text
38
+ replay.mp4 filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -1,6 +1,7 @@
1
  ---
2
  license: apache-2.0
3
  base_model: RenzKa/simlingo
 
4
  pipeline_tag: robotics
5
  language:
6
  - en
@@ -100,7 +101,7 @@ OUT_DIR=/tmp/personadrive_route00 \
100
  bash scripts/eval/test_one_lb1_route.sh
101
  ```
102
 
103
- - **Style:** set `PERSONADRIVE_ICL_PROFILE` to `fast` (aggressive), `default` (neutral), `slow` (conservative) or `normal` (no style; the PDM-lite expert index). `PERSONADRIVE_NUM_ICL=0` turns retrieval off.
104
  - **Retrieval data:** retrieval needs three things:
105
  - the retrieval head, included in the code repository;
106
  - the FAISS indices under `persona_drive_rag/`;
@@ -131,8 +132,8 @@ From conservative to aggressive, speed rises by 19.8 % and acceleration by 24.9
131
 
132
  - **Data:**
133
  - 2,848 PDM-lite routes;
134
- - 350 expert routes in the aggressive style (`fast/lb1_split`);
135
- - 350 expert routes in the conservative style (`slow/lb1_split`).
136
 
137
  Each frame is paired with its K = 2 context points, retrieved from the index of the matching style.
138
  - **Settings:** LoRA on the language model (r 32, α 64, dropout 0.1) with the vision encoder frozen. AdamW at lr 3e-5 (OneCycle, weight decay 0.1); batch 6 per GPU × 4 GPUs with DeepSpeed ZeRO-2 in bf16; 1 epoch.
 
1
  ---
2
  license: apache-2.0
3
  base_model: RenzKa/simlingo
4
+ base_model_relation: finetune
5
  pipeline_tag: robotics
6
  language:
7
  - en
 
101
  bash scripts/eval/test_one_lb1_route.sh
102
  ```
103
 
104
+ - **Style:** set `PERSONADRIVE_ICL_PROFILE` to `fast` (aggressive), `default` (neutral, normal driving), `slow` (conservative) or `normal` (no style; the index built from SimLingo's PDM-lite data). `PERSONADRIVE_NUM_ICL=0` turns retrieval off.
105
  - **Retrieval data:** retrieval needs three things:
106
  - the retrieval head, included in the code repository;
107
  - the FAISS indices under `persona_drive_rag/`;
 
132
 
133
  - **Data:**
134
  - 2,848 PDM-lite routes;
135
+ - 350 routes recorded in the aggressive style (`fast/lb1_split`);
136
+ - 350 routes recorded in the conservative style (`slow/lb1_split`).
137
 
138
  Each frame is paired with its K = 2 context points, retrieved from the index of the matching style.
139
  - **Settings:** LoRA on the language model (r 32, α 64, dropout 0.1) with the vision encoder frozen. AdamW at lr 3e-5 (OneCycle, weight decay 0.1); batch 6 per GPU × 4 GPUs with DeepSpeed ZeRO-2 in bf16; 1 epoch.
replay.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b446ce5e5cf51d6288552b359291564d93692539e4a61dbb8cf448d2899641d
3
+ size 14346690