---

license: other

base_model:

• HaoranLiu/SFT-4B-ScaleCUA-EvoCUA

language:

• en

pipeline_tag: image-text-to-text

library_name: transformers

tags:

• text-generation-inference
• llama-cpp

---

SFT-4B-ScaleCUA-EvoCUA-GGUF

> SFT-4B-ScaleCUA-EvoCUA is a Qwen3-VL-4B-Instruct checkpoint supervised-finetuned on single-teacher lite.scalecua trajectories sourced exclusively from EvoCUA-8B-20260105, isolating that teacher's individual contribution within the broader ScaleCUA campaign (EvoCUA is one of three teachers combined in the separate MixedOpen arm). Of 1,808 source rows, 437 were excluded via an exclude_reason/episode_return > 0.5 filter, leaving 865 usable trajectories (9,294 total steps, mean 10.74 per trajectory) — the smallest single-teacher dataset in the comparison set, versus 1,224 for the Qwen3.8-27B teacher arm, 1,075 for Qwen3.5-27B, and 1,539 for gpt-5.5. Training used an identical recipe across all arms — 3 epochs (648 steps), cosine LR from 5e-6 to 1e-6, global batch size 4 trajectories/step, bf16, seq length 4096, TP=2/DP=4 across 8x 80GB GPUs — completing in just under 4 hours with the lowest final training loss (0.044) among the single-teacher arms despite its smaller dataset. Two checkpoints are available (main at epoch 3/iter_647, and iter_431 at epoch 2); lite.osworld benchmark results (332 tasks, greedy decoding, max 30 steps) are listed as pending, with the base Qwen3-VL-4B-Instruct scoring 0.3062 mean and the Qwen3.8-27B and MixedOpen single/multi-teacher arms scoring 0.3644 and 0.3750 respectively for reference.

Model Files

File Name | Quant Type | File Size | File Link |

|-----------|------------|-----------|-----------|

| SFT-4B-ScaleCUA-EvoCUA.BF16.gguf | BF16 | 8.05 GB | Download |

| SFT-4B-ScaleCUA-EvoCUA.Q3_K_L.gguf | Q3_K_L | 2.24 GB | Download |

| SFT-4B-ScaleCUA-EvoCUA.Q3_K_M.gguf | Q3_K_M | 2.08 GB | Download |

| SFT-4B-ScaleCUA-EvoCUA.Q4_K_M.gguf | Q4_K_M | 2.5 GB | Download |

| SFT-4B-ScaleCUA-EvoCUA.Q4_K_S.gguf | Q4_K_S | 2.38 GB | Download |

| SFT-4B-ScaleCUA-EvoCUA.Q5_K_M.gguf | Q5_K_M | 2.89 GB | Download |

| SFT-4B-ScaleCUA-EvoCUA.Q5_K_S.gguf | Q5_K_S | 2.82 GB | Download |

| SFT-4B-ScaleCUA-EvoCUA.mmproj-bf16.gguf | mmproj-bf16 | 839 MB | Download |

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp