data-and-rewards
Prepare and validate TRL datasets, chat templates, multimodal messages, tool-call data, packing, and built-in reward functions before trainer construction.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- data-and-rewards
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Prepare and validate TRL datasets, chat templates, multimodal messages, tool-call data, packing, and built-in reward functions before trainer construction.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/api-reference.md4.5 KB
- 参考文档references/data-formats.md4.4 KB
- 参考文档references/reward-functions.md4.0 KB
- 参考文档references/troubleshooting.md4.5 KB
- 脚本scripts/validate_trl_dataset.py7.3 KB
03 SKILL.md 正文
TRL Data and Rewards
Use this sub-skill when a task is about shaping examples before they reach a TRL trainer, validating dataset columns, applying or cloning chat templates, preparing multimodal or tool-call messages, packing text datasets, or choosing built-in reward functions.
Route by task
- Dataset schema or conversion: start with Data Formats to identify standard vs conversational rows and task-specific columns.
- API behavior: use API Reference for
trl.data_utilsandtrl.chat_template_utilshelpers such asmaybe_apply_chat_template,maybe_extract_prompt,unpair_preference_dataset,pack_dataset, andparse_response. - Reward selection: use Reward Functions for GRPO/RLOO-compatible built-ins and the
None-means-skip behavior of math rewards. - Data validation: run
scripts/validate_trl_dataset.py --help, then validate local JSONL rows with--task sft,--task dpo,--task reward, or--task grpobefore trainer setup. - Failures: use Troubleshooting for missing columns, implicit prompts, invalid chat templates, multimodal image mismatches, tool-call parsing, and optional reward dependencies.
Boundaries
- This sub-skill covers dataset shape, preprocessing utilities, chat template preparation, multimodal/tool-call message shape, packing, and reward callable interfaces.
- For trainer class construction, callbacks, generation loops, or model/reference model setup, route to
core-training. - For YAML/CLI field names and launch recipes, route to
cli-and-configs. - For environment and tool-loop training behavior, route to
experimental-and-environments.
Safe workflow
- Classify the row type and format before changing code: standard text strings and conversational message lists have different valid columns.
- Normalize only when needed: use
maybe_*helpers when rows may already be standard, and explicit helpers when conversion is required. - Apply chat templates only to conversational examples and pass
toolsonly when the model/template supports tool calling. - Validate reward functions outside training with tiny in-memory completions or token-id lists; reward callables should accept extra trainer kwargs.
- Run the bundled validator on JSONL samples before constructing trainers so schema issues surface without downloads or training.
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).