ChronoReplay-8B

Model accompanying ChronoReplay: One Answer Interface Is Not Enough for Streaming Spatial Intelligence.

This is the frozen eight-GPU, seed-42 paper model, initialized from Qwen/Qwen3-VL-8B-Instruct revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b. It retains chronological replay, native categorical/numeric/temporal answer interfaces, and family-balanced supervision. The separate four-GPU replication is not substituted for this model.

Contents

Four full safetensors shards, their index, model configuration, tokenizer, image/video processor configuration, and chat template are included. No adapter merge is required. Optimizer states, training logs, and saved predictions are excluded. provenance.json binds the model to its training source; training_recipe.json records the training configuration.

Loading

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model_id = "mis2/ChronoReplay-8B"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

Use Transformers 4.57.3 and the accompanying ChronoReplay evaluation entry points for the frozen benchmark observation and decoding protocols. Loading these weights alone does not select an observation-access policy.

Training

• 39,561 serialized training records: 30,635 categorical, 2,545 numeric, and 6,381 supervised OST causal episodes.
• One epoch, global batch 32, learning rate 2e-6, cosine schedule, warmup ratio 0.03, weight decay 0.01.
• Maximum sequence length 32,768; maximum processed pixels per visual item 200,704; bfloat16; frozen visual encoder; trainable language model and merger.
• Seed 42, eight GPUs, microbatch two and accumulation two, GPU-only ZeRO-3.

Resources and terms

The ChronoReplay collection links this model and its training annotations. The implementation is available in the code repository. This release preserves the existing model weights. The base model and source datasets retain their original terms; this card does not relicense third-party materials. The upload is an artifact transfer, not a new training run or benchmark reevaluation.

See CITATION.cff for the author list and citation.