Nabidnur/bdb2027-combine-flm

Combine Motion FLM (V1)

A frame-level Foundation Motion Model for athletic movement, pretrained on NFL Big Data Bowl 2027 tracking data with three self-supervised objectives — masked-segment reconstruction (MAE), short-horizon forecasting, and InfoNCE identity — evaluated on the pre-registered MotionBench benchmark.

• GitHub (full stack, docs, benchmark): https://github.com/Hisernberg/nfl-bdb-2027
• Competition: NFL Big Data Bowl 2027 (deadline 2027-01-06)

Architecture

Component	Spec
Input	per-frame 9-dim kinematics (x, y, s, a, dis, o, dir + aligned velocity sin/cos), 10 Hz, 128-frame windows
Core	Mamba-2 d128×2 on GPU; GRU d128×2 fallback (this release: GRU, trained on CPU)
Heads	recon (masked frames), forecast (context→future offsets, residual on constant-velocity continuation), InfoNCE identity (proj dim 64)
Params	0.24 M

Training data

• Combine drills: 510 players, 463k frames (2023–2025 classes).
• Game tracking: 22.65M frames (2023–2025 seasons).
• 277,508 windows (7,014 combine + 270,494 game) in an fp16 memmap store; splits frozen by construction (combine 2023–24 + game REG 2023–24 → TRAIN; combine held-out drill classes + game 2025 wk1–9 → VAL; combine 2025 + game 2025 wk10–18 → TEST).

Results (real data, pre-registered MotionBench)

GPU mamba2 (12 epochs, T4×2): T1 PASS — ADE@10 0.249 yd vs constant-velocity 0.484 yd (−49% error); T2 PASS (7.61 yd @64 vs kNN 51.09); T7 PASS. Scoreboard: 3/6.

Gate	Result	Verdict
T2 long-horizon continuation	ADE@64 7.92 yd vs kNN 51.09 · ADE@128 8.37 yd vs 53.46	✅ PASS (6.4×)
T7 calibrated draft decision	conformal coverage 80.2% (76–84% target); net benefit 0.221 > 0.191 draft rule	✅ PASS
T1 short-horizon forecast	V0 0.586 → V1 0.497 yd vs CV 0.484 yd (residual-on-CV head)	at CV parity, −2.7%
T4 drill→game retrieval	recall@10 ≈ random at V0	❌→L3/L4
T6 draft-success probe	pAUC lift −0.14 at V0 (raw drills > embedding)	❌→L3/L4
T3 changepoint audit	30-attempt manual audit pending	⏳

T7 is the decision-relevant headline: the conformal-calibrated FLM beats the draft-capital rule on real data at V0.

Usage

import torch, sys
sys.path.insert(0, "path/to/repo")            # clone from GitHub
from src.evals.motionbench import load_model
model, ck = load_model("flm_best.pt", "cpu")  # 0.24M params
h, z = model(x, mask)                          # x [B,128,9] normalized, mask [B,128]
# z: motion embedding for retrieval / probes
# model.forecast(x, mask, horizon_eval=10): residual-CV future offsets

Input normalization uses TRAIN-split statistics shipped in the repo's data/proc_*/feat_stats.json (per-channel mean/std over valid frames).

Limitations

• V0/V1 is a 0.24M-param GRU trained ≤4 CPU epochs — it is a calibrated starting point, not a finished model; T1/T4/T6 gates are open.
• T3 requires human auditing (30 attempts) and is not_evaluated by design.
• Trained on NFL Combine and NFL game data (CC BY-NC 4.0) — non-commercial use.

Citation

@misc{combine_motion_flm_2026,
  title  = {Combine Motion FLM: A Frame-Level Foundation Model for Athletic Movement},
  author = {Hisernberg},
  year   = {2026},
  url    = {https://github.com/Hisernberg/nfl-bdb-2027}
}