pkq233

minicpmo45-matchtime-window-v5-lora-continue204

README

License: other

Intended Use

Generate concise English soccer commentary for a 30-second broadcast clip. This adapter is an exploratory research artifact. It is not intended for real-time match reporting or factual use without verification.

Training

Base revision: 4382fcae8a551b54d18f18462db974ff312aa7f3
Unique training examples: 300
Development examples: 100
LoRA targets: q_proj,v_proj
Trainable parameters: 3,833,856 (0.0441% of the loaded video-only path)
Vision tower and multimodal projector frozen
One pass over 96 examples, then one pass over 204 previously unseen examples
Learning rate: 1e-5
Effective batch size: 8
Precision: bfloat16
Hardware: Hugging Face Jobs, A100 80 GB

Evaluation

Evaluation used 100 frozen clips from 10 held-out matches, containing 112 reference events.

Table with columns: Metric, Base, LoRA-300
Metric	Base	LoRA-300
Micro event coverage	41.96%	44.64%
Contradiction rate	28.57%	24.11%
Net event score	13.39%	20.54%

Net event score assigns FULL=1, PARTIAL=0.5, MISS=0, and CONTRADICTION=-1. The LoRA-minus-Base match-cluster bootstrap 95% confidence interval was [+1.30, +12.50] percentage points. The ordinary event-coverage gain was directional; its confidence interval crossed zero.

Inference

The successful deterministic inference environment used transformers==4.51.0:

python
import torch
from peft import PeftModel
from transformers import AutoModel

base = AutoModel.from_pretrained(
    "openbmb/MiniCPM-o-4_5",
    revision="4382fcae8a551b54d18f18462db974ff312aa7f3",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    attn_implementation="sdpa",
    init_vision=True,
    init_audio=False,
    init_tts=False,
    low_cpu_mem_usage=True,
    device_map="cuda",
)
model = PeftModel.from_pretrained(
    base,
    "pkq233/minicpmo45-matchtime-window-v5-lora-continue204",
).merge_and_unload().eval()

Pass uniformly sampled PIL frames and the training prompt to model.chat with generate_audio=False, enable_thinking=False, and do_sample=False.

Limitations

Only 300 training examples were used.
The source annotations contain known timestamp and caption imperfections.
Most evaluation labels were produced by a text-only reference-grounded judge; only two clips were subsequently adjudicated against video.
This evaluates fixed-window caption generation, not proactive speaking time.
SoccerNet videos and MatchTime captions are not redistributed in this model repository. Users must obtain them from the official sources and comply with their terms.

License and Attribution

MiniCPM-o 4.5 is distributed under Apache-2.0. This adapter is published with license: other because the upstream dataset repository does not provide a clear standalone license for redistribution. No source media or annotations are included.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

pkq233

Model Tree

Base

openbmb/MiniCPM-o-4_5

Adapter

this model

Input Modalities

TextAudioImageVideo

Output Modalities

Text

Supported Functionality