Wing22/lvr-align-v2-1000

LVR Alignment v2 — step 1000

Complete model weights exported from the author's runs/align/v2/checkpoints/align-000-00001000.ckpt. This is the alignment checkpoint used to initialize LVR SFT v3 step 4000.

Code, training JSON, download scripts and evaluation tools: wmz422/lvr_pipeline_new.

Download and use

After installing the code repository and its dependencies:

python scripts/download_model.py --repo-id Wing22/lvr-align-v2-1000 --output models/lvr-align-v2-1000

from lvr.bundle import load_bundle
model, processor = load_bundle("models/lvr-align-v2-1000", device="cuda:0")

Use this checkpoint for known-latent alignment inference or SFT initialization. With the training data and recorded Qwen base model prepared as described in the code README:

ALIGN_CHECKPOINT_PATH=models/lvr-align-v2-1000 bash scripts/train_sft.sh

Browse all files, download standalone LAM, or get the SHA256 checksums.

This is a custom LVR bundle loaded by the accompanying repository, not a native Transformers AutoModel checkpoint. All three safetensors shards, the shard index, lvr_config.json and qwen/ resources are required. The bundle is self-contained for inference. It contains no optimizer or scheduler state and does not support exact mid-run optimizer resumption.

Training and reconstruction

• Source checkpoint: epoch 0, global step 1000; Lightning, approximately 22.49 GiB including optimizer state.
• Base: Qwen3-VL-4B-Instruct, recorded revision ebb281ec70b05090aa6165b016eac8ec08e71b17.
• Training: Qwen language model, tied LM head and two-layer latent projector; frozen Qwen vision encoder and LAM.
• Learning rates: Qwen 1e-6, projector 1e-4; effective batch size 64; BF16 mixed precision; alignment latent loss weight 0.
• LAM: 1024 model dimension, 32 latent dimension, four latent tokens, 16 encoder/decoder blocks, 16 heads.
• The source saves 402 trainable tensors. Export restores 1,199 frozen tensors from the recorded base/LAM files and one tied LM-head alias, producing 1,602 complete tensors.
• Token embeddings retain the recorded 151,672-token vocabulary, including the latent special tokens.
• Three BF16 safetensors shards total approximately 10.65 GiB. lam/lam.ckpt preserves original FP32 LAM parameters, approximately 3.32 GiB.

All exported tensors were saved and read back exactly. The export manifest includes hashes of the native checkpoint, source base-model weights, original LAM and exported files. Standalone LAM weights are checked against the bundled frozen LAM at the export dtype.

Verification scope

Local verification compares two known-latent alignment generations with the original checkpoint and unmodified source implementation, and exercises two VSTAR examples through the benchmark adapter. These are functionality checks; this alignment-stage model is not the final SFT model. No data downloads, training or full benchmark evaluation were run for publication. Details and records are in the code repository's results/alignment_step1000/ and docs/verification.md.

Recorded LAM pretraining includes some benchmark-derived images and auxiliary boxes, including answer-conditioned annotations. Refer to the code repository's data provenance; benchmark outputs do not represent a strictly held-out evaluation.

No separate release license has been specified for the author's code or exported weights. Upstream model, code and dataset terms remain applicable.