Buckets:
| Name | Size | Uploaded | Xet hash |
|---|---|---|---|
| .gitattributes | 1.79 kB xet | cc74b6dc | |
| LFM2.5-VL-450M-IN-CAR-SCENE-SEATS-v4-F16.gguf | 846 MB xet | ad78ce6b | |
| LFM2.5-VL-450M-IN-CAR-SCENE-SEATS-v4-Q4_0.gguf | 257 MB xet | c9ab0e67 | |
| README.md | 3.16 kB xet | 292dff72 | |
| chat_template.jinja | 3.84 kB xet | 357a20ec | |
| config.json | 2.56 kB xet | a9b0c1c9 | |
| generation_config.json | 148 Bytes xet | 35f1f71d | |
| mmproj-F16.gguf | 189 MB xet | 7cc30f54 | |
| mmproj-Q4_0.gguf | 55 MB xet | 0f5950fd | |
| model.safetensors | 1.03 GB xet | 4130b2b7 | |
| processor_config.json | 827 Bytes xet | e57ff681 | |
| scene_description_flat_reduced.gbnf | 1.07 kB xet | fdc93a25 | |
| seats_flat.gbnf | 219 Bytes xet | aedb506e | |
| tokenizer.json | 4.73 MB xet | e061c89e | |
| tokenizer_config.json | 903 Bytes xet | c140a767 |
LFM2.5-VL-450M — In-Car SCENE+SEATS v4 (flat, 64 image tokens)
Full fine-tune of LiquidAI/LFM2.5-VL-450M on the scene and seats tasks only (no pointing). Task selected by a short task tag in the user turn; each emits a compact flat (pipe-delimited, key-less) line. Quantized to Q4_0 and F16.
Internal work-in-progress (v4 scene+seats). Kept private.
Sister checkpoint trained on all three tasks: LFM2.5-VL-450M-IN-CAR-COMBINED-v4-GGUF.
Task routing
The user message is an image + one tag (plain text):
| tag | task | output line |
|---|---|---|
scene |
cabin scene description | lighting|occupant_count|driver_present|driver_distraction|driver_face|hazards|notable_objects|summary |
seats |
per-seat occupancy | driver|front_passenger|rear_left|rear_middle|rear_right (blank = empty seat) |
Fields separated by |, list items by ,. Face boxes are xmin,ymin,xmax,ymax
integers in [0,1000]. Full contract: COMBINED_SPEC.md in the training repo.
# user: <image> + "scene"
day|1|true|phone_call|239,298,352,571|||An elderly man drives while on the phone.
# user: <image> + "seats" (driver adult, rear_right child, rest empty)
adult||||child
Files
| file | what |
|---|---|
LFM2.5-VL-450M-IN-CAR-SCENE-SEATS-v4-Q4_0.gguf |
text model, Q4_0 (~210 MB) |
LFM2.5-VL-450M-IN-CAR-SCENE-SEATS-v4-F16.gguf |
text model, F16 (~679 MB) |
mmproj-Q4_0.gguf |
vision projector, Q4_0 (~53 MB) |
mmproj-F16.gguf |
vision projector, F16 (~181 MB) |
scene_description_flat_reduced.gbnf |
grammar for the scene contract (8 fields) |
seats_flat.gbnf |
grammar for the seats contract |
Image preprocessing (required)
Resize each frame so ceil(w/32)*ceil(h/32) ≲ 64 at source aspect ratio, then
pass --image-min-tokens 32 --image-max-tokens 64.
Run (llama.cpp)
llama-server \
--model LFM2.5-VL-450M-IN-CAR-SCENE-SEATS-v4-Q4_0.gguf \
--mmproj mmproj-Q4_0.gguf \
--host 127.0.0.1 --port 8080 --ctx-size 2048 \
--image-min-tokens 32 --image-max-tokens 64 \
--temp 0.1 --min-p 0.15 --repeat-penalty 1.05 --seed 42 --jinja
Sampling: temperature=0.1, min_p=0.15, repetition_penalty=1.05.
Training
| corpus | combined_v4_scene_seats — 93,046 train / 4,902 eval (scene + seats, 1:1) |
| data | hybrid-gold synthetic cabins from combined_v4_new, wheel on image-right |
| recipe | full FT, 2 epochs, 4×MI325X, per-device 256, global 1024, lr=1e-4 cosine, warmup 0.1 |
| vision | max_image_tokens=64, do_image_splitting=false, assistant_only_loss=true |
| checkpoint | job 1952597, step 182 (end of epoch 2) |
Evaluation
Not reported for this build.
- Total size
- 2.38 GB
- Files
- 15
- Last updated
- Sep 22
- Pre-warmed CDN
- US EU US EU