LEAP Checkpoints (leap-ckpts): Full-Parameter SFT Trajectory on Qwen3-VL-2B-Instruct

Model Summary

EvanSirius/leap-ckpts hosts the full-parameter supervised fine-tuning (Full SFT) checkpoint trajectory (checkpoint-200 through checkpoint-7000, saved every 200 optimization steps) of Qwen/Qwen3-VL-2B-Instruct for "Structured Supervision Targets as a Knowledge Interface for Reliable Parser-Gated Plan Generation in Compact Vision-Language Models".

All checkpoints are trained on 151,230 single-frame ordered-step instances (EvanSirius/leap-agibot-processed) using BF16 precision, SDPA attention, response-only loss masking, cosine learning rate decay (peak LR $1.5\times 10^{-5}$, 800-step linear warmup), and effective batch size 36 (6Γ—6 gradient accumulation).

VLABench Evaluation Summary (checkpoint-5000 Canonical Release)

Evaluated across 4,500 episodes spanning all six VLABench capability dimensions under deterministic greedy decoding:

Evaluation Regime Qwen3-VL-2B (Base) Same-Domain Control ($\mathcal{T}_{\mathrm{ctrl}}$) LoRA r=16 (5k) Full SFT (checkpoint-5000)
Strict Parser-Gated Macro Score ($\psi_{\mathrm{raw}}$) 25.10 25.15 28.22 28.38 ([28.01, 28.75])
Stage 1 Syntax Parse Survival Rate 99.18% β€” 87.75% 96.00%
Stage 2 Entity Type-Normalized Macro Score ($\psi_{\mathrm{norm}}$) 35.66 β€” 33.58 38.90
Stage 2 Skill-Sequence Match Rate 62.52% β€” 60.12% 70.68%
Stage 2 Joint Step Match Rate 18.13% β€” 18.31% 22.05%
Stage 2 Exact Plan Match Rate 4.33% β€” 2.41% 6.60%

Usage

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "EvanSirius/leap-ckpts"
subfolder = "checkpoint-5000"

processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-2B-Instruct")
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    subfolder=subfolder,
    torch_dtype=torch.bfloat16,
    device_map="auto",
).eval()

Related Resources

Citation

@article{gao2026structured,
  title={Structured Supervision Targets as a Knowledge Interface for Reliable Parser-Gated Plan Generation in Compact Vision-Language Models},
  author={Shenzhou Gao and Peijian Zhang},
  year={2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for EvanSirius/leap-ckpts

Finetuned
(316)
this model