One-line identity: Stage 1 NTP SFT fine-tune of Qwen/Qwen2.5-VL-7B-Instruct on the ohjoonhee/visual-cot-50k-poc dataset (Monet-SFT-125K Visual_CoT subset). Trains the model to emit <observation> and other Monet special tokens fluently before Stage 2's alignment objective layers in latent slots. Baseline reference for downstream Stage 2 / Pivot A runs.

Recipe

Stage: 1 (NTP SFT; no alignment, no latent slots)
Base model: Qwen/Qwen2.5-VL-7B-Instruct
Init checkpoint: (none)
Dataset: ohjoonhee/visual-cot-50k-poc (Monet-SFT-125K Visual_CoT subset, eval-200 excluded)
Hardware: 4× H100 80GB, DeepSpeed ZeRO-2 + CPU optim offload, bf16
lr: 1e-05
weight_decay: 0.0
warmup_steps: 100
max_steps: 1100
grad_accum_steps: 16
max_pixels: 784000

Notes

Pure NTP SFT — no Monet Stage 2 alignment loss, no latent-mode forward. The Monet special tokens (<observation>, <abs_vis_token>, etc.) ARE registered in the tokenizer and embedded so the model learns to produce them, but the architectural latent-slot mechanism is unused at this stage.

This revision (`step-1100`)

Last logged training row: step=1100, elapsed=38112s

Notes

Stage 1 NTP SFT baseline; final ckpt at step 1100, ntp_loss=0.288

Other revisions: see the revisions dropdown on this page.

How to load

python

from transformers import AutoModelForVision2Seq, AutoProcessor
m = AutoModelForVision2Seq.from_pretrained(
    "ohjoonhee/vlatents-qwen25vl7b-stage1-sft-v1", revision="step-1100", torch_dtype="bfloat16")
p = AutoProcessor.from_pretrained("ohjoonhee/vlatents-qwen25vl7b-stage1-sft-v1", revision="step-1100")

Limitations

Research checkpoint, eval-only. Final checkpoint at step 1100/1100. Not for production.

Card generated 2026-05-15 from training_log.jsonl + the run's training config.

ohjoonhee/vlatents-qwen25vl7b-stage1-sft-v1

vlatents-qwen25vl7b-stage1-sft-v1

Recipe

Notes

This revision (step-1100)

Notes

How to load

Limitations

ohjoonhee/
vlatents-qwen25vl7b-stage1-sft-v1

This revision (`step-1100`)