modeling
Choose, instantiate, fit, score, and recommend with Microsoft Recommenders model families across verified CPU and optional Spark, GPU, and experimental backends.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- modeling
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Choose, instantiate, fit, score, and recommend with Microsoft Recommenders model families across verified CPU and optional Spark, GPU, and experimental backends.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/api-reference.md10.9 KB
- 参考文档references/model-overview.md9.7 KB
- 参考文档references/troubleshooting.md9.0 KB
- 参考文档references/workflows.md11.3 KB
- 脚本scripts/sar_tiny_smoke.py5.9 KB
- 脚本scripts/tfidf_tiny_smoke.py6.0 KB
03 SKILL.md 正文
Modeling
Use this sub-skill when a user asks which Recommenders model family to use, how to instantiate and fit a model, how to produce predictions or top-k recommendations, or how optional Spark/deep-learning/content-based/experimental model paths differ.
Backend truth for this skill
- Verified in the base CPU scope: package import, SAR, TF-IDF utilities, Cornac prediction helpers, and LightGBM helper imports/signatures.
- Optional and not verified in this CPU scope: Spark ALS/SARplus/Spark LightGBM, TensorFlow/PyTorch deep-learning models, NewsRec/DeepRec notebooks, cloud workflows, and experimental packages such as Surprise, LightFM, Vowpal Wabbit, xLearn, GeoIMC, and RLRMC dependencies.
- Do not present a visible GPU, Spark, or cloud host as sufficient. The matching optional Python packages, system runtime, data files, and credentials must also be present.
Start here
- For algorithm choice across SAR, TF-IDF, Cornac, LightGBM, Spark ALS, NewsRec, NCF, Wide&Deep, VAE, RBM, SASRec/SSEPT, and experimental families, read model-overview.md.
- For constructors, methods, dataframe contracts, and return columns, read api-reference.md.
- For runnable recipes, use workflows.md:
- CPU SAR collaborative filtering with
SAR.fit,SAR.predict, andSAR.recommend_k_items. - CPU TF-IDF content-based recommendation with
TfidfRecommender. - Cornac scoring/ranking helpers.
- LightGBM feature encoding helper usage.
- Optional Spark, deep-learning, and news/sequential model preparation checklists.
- CPU SAR collaborative filtering with
- For symptoms and fixes, read troubleshooting.md.
- To prove a local install can run the two lightweight CPU model paths, run:
python sub-skills/modeling/scripts/sar_tiny_smoke.py --top-k 2
python sub-skills/modeling/scripts/tfidf_tiny_smoke.py --top-k 1
Run the commands from the root of this generated skill tree, or adapt the script paths to the location where the skill was installed.
Route elsewhere
- Data loading, schema validation, de-duplication before training, filtering, negative sampling, sparse matrix conversion, and train/test splitting belong in data-preparation.
- Metric calculation after predictions or top-k output belongs in evaluation.
- Hyperparameter sweeps, benchmark harnesses, NNI/AzureML/Databricks/AKS, service sizing, and long-running operationalization belong in operations-and-tuning.
Working rules
- Identify the available data signal first: interaction ratings/clicks, item text/content, tabular ad features, sequence/session logs, news/MIND files, or Spark-scale data.
- Identify the permitted backend: base CPU only, Spark, GPU/deep-learning, experimental native packages, or cloud services.
- Pick the smallest verified path that matches the data. Prefer SAR for a lightweight pandas interaction baseline and TF-IDF for text-only item similarity before escalating to optional models.
- Keep
userID,itemID,rating,timestamp, andpredictioncolumn names unless the user explicitly uses custom names and passes matchingcol_*parameters. - When producing recommendations for evaluation, ensure top-k output has one row per
(userID, itemID)candidate and apredictionscore, then route metrics to the evaluation sub-skill.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).