[iammosespaulr]

Adds PDM Femto, Pico, and Nano from
Flashmind UAE, evaluated through the
flashpdm Python client.

The client handles batching, concurrency, and forecast recovery; the wrapper
maps fev's inputs and prediction format.

• models/flashpdm/model.py and requirements.txt: wrapper and pinned client dependency.
• benchmarks/fev_bench/results/pdm-{femto,pico,nano}.csv: results on all 100 tasks.
• benchmarks/fev_bench/models.yaml: three model entries.

Model	Parameters	SQL skill ↑	MASE skill ↑
PDM-Femto	6.18M	47.90%	36.60%
PDM-Pico	32.2M	49.09%	37.78%
PDM-Nano	119M	50.33%	39.21%

Skill scores are relative to Seasonal Naive, computed with fev's official
leaderboard aggregation over all 100 tasks.

Zero-shot, served through a closed API. No test-data leakage.
Targets are forecast jointly, including when known-future covariates are supplied. Past and known-future covariates are used; string categories use
per-series target encoding from observed history only. Static columns are unused.
Context is capped at 16,336 steps. Forecasts use flip averaging, nonnegative
projection and integer-count rounding determined from observed targets.

Reproduce one model:

export FLASHMIND_API_KEY=your_evaluator_key
python models/evaluate.py -m flashpdm -b benchmarks/fev_bench/tasks.yaml \
  -n PDM-Nano -k '{"model":"pdm-nano"}'

For Femto or Pico, change the model name and display name accordingly. Model
revisions are pinned, and dependencies are installed by the runner. Inference uses the
Flashmind API; no local GPU is needed.
Evaluator credentials with sufficient quota are available to maintainers through
Flashmind.

[iammosespaulr]

Hi @shchur! Our submission uses the Flashmind API. We’ve prepared an evaluator account and API key using shchuro @ amazon.com with enough quota to reproduce all three models and run additional evaluations.

We’re happy to email you the key, or you can sign in at https://www.flashmind.com/dashboard/api-keys and click Show key beside fev-bench reproduction to copy it. Let us know which you prefer, happy to help!

[shchur]

Hi @iammosespaulr, thank you for the contribution. Can you please set hidden: true for the smaller model sizes, so that only Nano shows up on the leaderboard? We typically prefer to add just 1 model to the main leaderboard to avoid having multiple highly correlated models that dilute the win rates.