Downloads · 30 days
8
14% of all-time downloads
lrmbbj/cl_0.8B_t0.07_q8_d256_nt16x16_fr4-16_do0.1_rc0.1_jt0.25_s42-1000step
cl_0.8B_t0.07_q8_d256_nt16x16_fr4-16_do0.1_rc0.1_jt0.25_s42-1000step is a feature extraction model from lrmbbj. Use it when you need embeddings to search or compare text. It is set up for transformers.
MCE(Multimodal Context Encoder,VLM backbone + 连续预测头)的打包导出, 自定义代码加载(trustremotecode=True),部署端不依赖训练仓库。
Downloads · 30 days
8
14% of all-time downloads
All-time downloads
58
Public
Parameters
856M
3.4 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors1.7 GB · 99%
From the Hugging Face model README
MCE(Multimodal Context Encoder,VLM backbone + 连续预测头)的打包导出,
自定义代码加载(trust_remote_code=True),部署端不依赖训练仓库。
import torch
from transformers import AutoModel
model = AutoModel.from_pretrained(
"lrmbbj/cl_0.8B_t0.07_q8_d256_nt16x16_fr4-16_do0.1_rc0.1_jt0.25_s42-1000step", trust_remote_code=True, dtype=torch.bfloat16
).eval().cuda()
# 纯上下文编码:任意上下文模态子集 -> 动态嵌入 [B, D]
z = model.encode_context(
task_description="pick up the cube", # 可选
goal_image=None, # 可选,单帧图像
robot_demo=demo_video, # 可选,[T,C,H,W] uint8/float
)
# 同时拿对比嵌入与原始 query token 隐状态:
out = model.encode_with_tokens(
task_description="pick up the cube",
robot_demo=demo_video,
)
out["embedding"] # [B, D],与 encode_context 一致(归一化的对比嵌入)
out["tokens"] # [B, Q, H],Q 个可学习 query token 的隐状态(未投影/池化)
torch(>= 2.1)transformers:需能加载 Qwen/Qwen3.5-0.8B 的版本(以基座 repo
config.json 里的 transformers_version 为下限;本 checkpoint 导出
时使用的是 transformers 5.4,建议部署端保持一致或更新)omegaconf、pillow、numpy图像输入支持 Tensor / numpy / PIL(uint8 [0,255] 或 float [0,1]),
内部自动 resize 到训练时的 image_size。