Dataset Mixtures

Modern VLA training often mixes demonstrations from different robots, tasks, and source formats. kinodb gives you two paths:

• kino merge creates one physical .kdb file.
• kino mix and from_mixture() create weighted virtual mixtures.

Physical Merge

Use merge when you want a single file to ship, archive, upload, or train sequentially.

kino merge lift.kdb pusht.kdb aloha.kdb --output combined.kdb

Filter while merging:

kino merge lift.kdb pusht.kdb \
  --output successful.kdb \
  --filter "success = true"

This reads each input episode, applies the optional KQL filter, then writes matching episodes to a new database.

Weighted Mixture CLI

Use mix when you want training-time sampling proportions.

kino mix \
  --source bridge.kdb:0.4 \
  --source aloha.kdb:0.3 \
  --source libero.kdb:0.3

Sample a distribution:

kino mix \
  --source bridge.kdb:0.4 \
  --source aloha.kdb:0.3 \
  --source libero.kdb:0.3 \
  --sample 1000 \
  --seed 42

Weights are relative. 4:3:3 and 0.4:0.3:0.3 are equivalent.

Python Mixtures

from kinodb.torch import from_mixture
from torch.utils.data import DataLoader

dataset = from_mixture(
    {
        "bridge.kdb": 0.4,
        "aloha.kdb": 0.3,
        "libero.kdb": 0.3,
    },
    seed=42,
    image_size=(224, 224),
)

loader = DataLoader(dataset, batch_size=8)

Each source can come from a different original format. The training code only sees .kdb.

Rust API

use kinodb_core::Mixture;

let mut mix = Mixture::builder()
    .add("bridge.kdb", 0.4)
    .add("aloha.kdb", 0.3)
    .add("libero.kdb", 0.3)
    .seed(42)
    .build()?;

let episode = mix.sample()?;
let global_episode = mix.read_global(10)?;
let order = mix.weighted_epoch(1000);

Mixed Schema Reality

Different datasets can have different action and state dimensions. The experiment history hit this directly when mixing PushT (action_dim = 2) with ALOHA (action_dim = 14): raw torch.stack fails unless the collate function pads or batches by schema.

Common strategies:

Strategy	When to use
Pad state/action vectors to max dimension	One model with source-aware masks
Bucket by schema	Multi-embodiment training with separate heads
Train separate adapters	Different embodiments have genuinely different action spaces
Physical merge only same-schema data	Simplest archival/distribution path

kinodb preserves dimensions and metadata; it does not hide schema differences from your model.

When To Merge vs Mix

Use case	Pick
Publish one converted dataset	kino merge
Build a filtered release split	kino merge --filter
Match OpenVLA-style source proportions	kino mix or from_mixture()
Change ratios between runs	virtual mixture
Maximize sequential reads from one file	physical merge