modeling-and-evaluation
Use for H2O LLM Studio model wrappers, losses, metrics, evaluation outputs, inference routing, plots, and AI-judge metric behavior.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- modeling-and-evaluation
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use for H2O LLM Studio model wrappers, losses, metrics, evaluation outputs, inference routing, plots, and AI-judge metric behavior.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/evaluation-workflows.md8.2 KB
- 参考文档references/model-and-metric-reference.md9.7 KB
- 参考文档references/troubleshooting.md7.7 KB
- 脚本scripts/inspect_problem_type.py13.7 KB
03 SKILL.md 正文
H2O LLM Studio modeling and evaluation
Use this sub-skill when the task is about how H2O LLM Studio maps a problem type to model wrappers, losses, metrics, inference behavior, generated prediction outputs, validation plots, or AI-judge metrics.
Do not use this sub-skill for:
- creating or repairing experiment YAML/data schemas: use
configuration-and-data; - launching, scheduling, or debugging training runs: use
training-and-experiments; - interactive chat, model-card export, or Hugging Face publishing: use
export-and-prompt.
Operating workflow
- Identify the
problem_typefirst. Valid modeling/evaluation problem types are:text_causal_language_modeling,text_sequence_to_sequence_modeling,text_dpo_modeling,text_causal_classification_modeling, andtext_causal_regression_modeling. - Load references/model-and-metric-reference.md for class mappings, forward/generate contracts, losses, metric direction, and expected result keys.
- Load references/evaluation-workflows.md for
run_inference,run_eval, prediction files, plot files, generation-vs-forward routing, and GPT/MT-Bench judge behavior. - For unsupported metrics, NaNs, empty outputs, OpenAI/Azure endpoint issues, missing
prompts/, or classification/regression shape errors, load references/troubleshooting.md. - Use scripts/inspect_problem_type.py for a safe local inspection that lists the configured model/loss/metric/plot classes without instantiating or downloading a model.
Quick commands
python scripts/inspect_problem_type.py --problem-type text_causal_classification_modeling --list-metrics
python scripts/inspect_problem_type.py --problem-type all --json
python scripts/inspect_problem_type.py --problem-type text_dpo_modeling --verify-imports
The inspection script is safe by default: it prints static and import-level metadata only. It does not start training, call an AI judge, load Hugging Face weights, or create model instances.
Decision rules
- For generative validation metrics other than
Perplexity, causal LM, sequence-to-sequence, and DPO use generation;Perplexityuses a forward pass. - Classification and regression are non-generation tasks. They always use a forward pass and postprocess logits or regression head outputs into predictions before metrics are computed.
GPTmetrics call an OpenAI-compatible Chat Completions endpoint and can incur network cost. Require endpoint, credential, and budget confirmation before running; use static reasoning or mocked tests when possible.- Validation artifacts are written by the training/evaluation workflow, not by the model classes themselves. Expect raw prediction pickle, prediction CSV, and parquet plot data when evaluation reaches rank 0 successfully.
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).