liantian/fastfill-v3

FastFill v3 — single-call object-conditioned room layout SFT data (private, internal research)

Built 2026-09-25 by fastfill.build (code snapshot in code/ of this repo). One training sample = one furnished room from one of 16 public indoor-scene datasets, turned into a chat turn: the user message holds the room (floor polygon, optional room type and ceiling height, fixed boxes already in the room, the list of objects to place with their local sizes, optional constraints) and the assistant message holds every object's position, yaw and support. See fastfill/README.md in the code repository for the full specification (sections 2 and 5).

Files

path	what
code/fastfill/	the pipeline package (build, train, evaluate, serve, interface, adapters); code/fastfill/FastFill_v1_定稿.md is the scope (what the model does, the data rules, training and evaluation), code/fastfill/README.md has every detail and number
code/scripts/merge_lora.py	merge the LoRA adapter into the base model
code/requirements-cu128.txt	pinned server environment for CUDA 12.8
ir/	the 18 sources in the unified intermediate format the build reads (metres, Z-up, bottom-centre boxes, yaw to the semantic front; one room per line), byte-identical to the files hashed in v3/MANIFEST.json; rebuild with python -m fastfill.build --ir ../ir --out ../rebuilt/v3 --sources <MANIFEST args.sources> (add --constraint_frac 0.3 for v3.1)
v3/	(R, G, O) -> L, no constraint inputs: train.jsonl, dev.jsonl, test.jsonl (chat rows: uid, source, flags, messages), dev_rooms.jsonl, test_rooms.jsonl, dev_constrained_rooms.jsonl, test_constrained_rooms.jsonl (room IR for fastfill.evaluate), stats.json, MANIFEST.json, QA.json
v3.1/	same rooms; 30% of the train rows carry 1-4 constraints (R, G, O, C) -> L; dev* / test* byte-identical to v3/

flags per row are quality labels (oob_objects, hidden_obstacle, overlapping_furniture, small_area_m2, room_filling_object, single_object, z_snapped, tilted, no_front_fixed, fixed_collision), not filters: every room whose training target could be built is included. train.py --exclude_flags <flag...> leaves out rows with those flags and --source_weight SOURCE=w uses each row of a source w times on average (0.5 halves it, 2 doubles it); the default leaves out rows flagged oob_objects, overlapping_furniture or fixed_collision (reference answers the server would refuse) and uses every other row once, in the data's own proportions; --exclude_flags with no value trains on all rows.

Server: install, download, train (CUDA 12.8)

All commands run from the downloaded code/ directory (it contains the fastfill package).

# 0. environment (skip if the node03 fastfill env with torch 2.10 / vllm 0.19.0 / transformers 4.57.6 already exists)
conda create -n fastfill python=3.11 -y && conda activate fastfill
pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cu128

# 1. code + data (private repo: log in once with a read token from https://huggingface.co/settings/tokens)
pip install -U "huggingface_hub>=0.36" && hf auth login
hf download liantian/fastfill-v3 --repo-type dataset --local-dir ~/fastfill-v3
cd ~/fastfill-v3/code && pip install -r requirements-cu128.txt
python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available(), torch.cuda.device_count())"
python -m fastfill.scene && python -m fastfill.validate && python -m fastfill.interface      # self-checks

# 2. base model
hf download Qwen/Qwen3-8B --local-dir ~/models/Qwen3-8B

# 3. look at the data with the real tokenizer (no GPU): token / object percentiles, samples over 40960
python -m fastfill.train --model ~/models/Qwen3-8B --data ~/fastfill-v3/v3 --out /tmp/x --dry_run 2>&1 | tee dry_run.txt

# 4. measure GPU memory on ONE card before choosing the batch size (one training step per length, padded like training)
CUDA_VISIBLE_DEVICES=0 python -m fastfill.train --model ~/models/Qwen3-8B --data ~/fastfill-v3/v3 --out /tmp/x --bs 1 --mem_test 8192,16384,32768,max
CUDA_VISIBLE_DEVICES=0 python -m fastfill.train --model ~/models/Qwen3-8B --data ~/fastfill-v3/v3 --out /tmp/x --bs 2 --mem_test 8192,16384,max
#    pick the largest --bs whose "max" row fits with a few GB to spare; keep bs x grad_accum x n_gpus = 32

# 5. train (8 GPUs; bs 1 -> grad_accum 4, bs 2 -> 2, bs 4 -> 1)
PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True torchrun --nproc_per_node 8 -m fastfill.train \
    --model ~/models/Qwen3-8B --data ~/fastfill-v3/v3 --out outputs/ff-v3 \
    --bs 1 --grad_accum 4 --lr 1e-4 --epochs 2 --save_steps 500
#    constrained variant: --data ~/fastfill-v3/v3.1 --out outputs/ff-v3.1 (same settings)
#    fewer GPUs: CUDA_VISIBLE_DEVICES=... and --nproc_per_node N, grad_accum = 32 / (bs x N)
#    interrupted: rerun the same command with --resume
#    if the step-500 evaluation runs out of memory on the longest dev rows: add --eval_max_len 32768

# 6. merge the adapter, then evaluate (vLLM)
python scripts/merge_lora.py --base_model_path ~/models/Qwen3-8B --lora_path outputs/ff-v3/final --output_path outputs/ff-v3-merged
python -m fastfill.evaluate --gt-only --rooms ~/fastfill-v3/v3/test_rooms.jsonl --out eval/gt
python -m fastfill.evaluate --model outputs/ff-v3-merged --rooms ~/fastfill-v3/v3/test_rooms.jsonl --out eval/ff-v3
python -m fastfill.evaluate --model outputs/ff-v3-merged --rooms ~/fastfill-v3/v3/test_constrained_rooms.jsonl --out eval/ff-v3-cons

# 7. serve for EmbodiedGen's FastFillBackend
python -m fastfill.serve --model outputs/ff-v3-merged --port 8001

Numbers

(from code/fastfill/README.md section 3, in Chinese)

train 144,140 / dev 8,167 / test 8,647（另有 dev_constrained、test_constrained：每间 1–4 条在参考布局上成立的约束）。

• v3 = 无约束基线 (R,G,O)→L；v3.1 = 同一批房间，其中 42,743 条训练样本带 1–4 条约束 (R,G,O,C)→L。两版 dev/test 逐字节相同（True）；tools/ablation_check.py：去掉 constraints 后两版训练样本差异 0 条。
• QA.json：泄漏（train 与 dev/test 之间 group、别名、家具指纹、布局指纹、held-out 源）全部为 0；答案恰好摆了输入里的全部物体 144,140/144,140；固定件出现在答案里 0 次；约束在参考布局上成立 18662/18662（dev）、19785/19785（test）。
• 参考布局本身的合法率（dev）：93.2%；不合法的是带 oob_objects 标记、原来会被拒收的房间。模型指标一律和同一批房间的 GT 值对比。
• token（Qwen3 分词器，train.py --dry_run）：训练集共 1.86 亿 token；单条 P50 896、P90 2,226、P95 3,440、P99 6,142、最长 135,965；≤4096 的 96.6%，≤8192 的 99.69%。每间待摆物体 P50 6、P90 28、P95 46、P99 84、最长 2,080；固定件 P50 2、P90 7、P95 8、P99 12、最长 70。
• 超过 40960 的训练样本 9 条：interiorgs:0044_839926::room_0（InteriorGS，2080 件，135,965 token）；interiorgs:0045_839925::room_0（InteriorGS，680 件，43,739 token）；interiorgs:0134_840039::room_0（InteriorGS，1274 件，80,252 token）；interiorgs:0141_840177::room_0（InteriorGS，1575 件，99,434 token）；interiorgs:0175_840149::room_0（InteriorGS，1229 件，67,984 token）；interiorgs:0179_840114::room_0（InteriorGS，1164 件，75,057 token）；interiorgs:0394_840186::room_0（InteriorGS，983 件，57,821 token）；interiorgs:0408_840112::room_0（InteriorGS，904 件，57,764 token）；MansionWorld::mansionworld/public_entertainment_4f_200_fp001#0/F2_event_hall（MansionWorld，628 件，44,392 token）。train.py 加载时整条跳过、不截断。dev 最长 16,584、test 最长 27,156。
• 固定件共 477,666 个，分布在 135,205 间写出的房间里（占 84%）：structure 441,532、generic 22,040、floating 8,800、no_front 5,128、wall 166。
• v3 修掉了全量审核在 v2 上查出的数据问题（MansionWorld 桌面物件、InteriorGS 门窗、倾斜框、3D-FRONT 精确边界和补门、厚墙门窗、误拒房间、物体描述、稳定划分），正式训练用 v3；v2 / v2.1 只用于跑通流程，test 集不同，数字不可互比。

各源房间数（train / dev / test）、单间最多待摆物体、固定件数、带固定件的房间数（都按写出的房间计）：

训练源	train / dev / test	最多物体	固定件	带固定件的房间
SpatialLM	42,068 / 2,239 / 2,461	43	156,103	46,334
MansionWorld	19,396 / 1,153 / 1,313	628	86,316	21,762
IL3D_3dfront	18,054 / 939 / 872	27	39,427	13,567
Structured3D	12,700 / 793 / 747	181	73,356	14,238
InternScenes_gen	12,552 / 828 / 660	140	59,378	14,040
SAGE-10k	8,977 / 503 / 489	189	12,991	9,969
OptiScene_holodeck	8,946 / 497 / 503	21	5	5
IL3D_synthetic	6,277 / 349 / 316	20	1,096	999
InteriorGS	4,867 / 268 / 311	2080	23,508	5,347
InternScenes_arkit	4,528 / 280 / 229	63	9,486	4,012
HSSD200	1,876 / 86 / 138	623	3,529	1,345
InternScenes_scannet	1,309 / 80 / 104	90	4,194	1,257
InternScenes_mp3d	1,258 / 89 / 18	112	5,170	1,247
InternScenes_3rscan	1,200 / 61 / 70	160	2,588	915
MultiScan	114 / 2 / 6	80	344	101
Scan2CAD	18 / 0 / 1	32	0	0
SceneSmith（只评测）	0 / 0 / 364	152	26	25
SpatialGen（只评测）	0 / 0 / 45	25	149	42

扫描 188,670 间。硬过滤拒收（间）：object_count 19,606、boundary 722、boundary_shape 142、incomplete_objects 113。

• object_count：没有一件可摆的物体（只有墙上、顶上的东西或没有名字的箱子），没有训练答案。
• boundary / boundary_shape：没有边界或边界多边形无效。incomplete_objects：有物体没有包围盒（HSSD 的可动柜门、冰箱等），会成为看不见的障碍。bad_numbers：数值非法，或要摆的物体尺寸 ≤0。
• 另外，每个 split 里完全相同的房间只留一份（duplicate_layout 6,137 间），跨源同一场景只留一个源（dedup_dropped 996 间）。

质量标记（写出的房间里，按标记计）：tilted 14,073、single_object 13,344、fixed_collision 11,918、oob_objects 10,959、z_snapped 5,059、no_front_fixed 3,480、hidden_m2 2,443、room_filling_object 2,193、overlapping_furniture 1,989、hidden_obstacle 1,257、small_area_m2 59。

Licenses

Derived from SpatialLM (CC BY-NC 4.0), IL3D / 3D-FRONT (3D-FRONT terms: research only, no redistribution), HSSD-200 (CC BY-NC 4.0), InternScenes (CC BY-NC-SA 4.0 + upstream scan terms), MultiScan (CC BY-NC 4.0), Structured3D (non-commercial research, no redistribution), InteriorGS (terms of use), MansionWorld (CC BY 4.0, non-commercial gate), SAGE-10k and OptiScene/Holodeck (Apache-2.0), Scan2CAD (ScanNet terms), SpatialGen (CC BY-NC 4.0), SceneSmith (Apache-2.0). This repository is PRIVATE and exists only to move the built files to the training server; it must not be made public or shared outside the research group, and models trained on it are internal, non-commercial only.