core-training
Choose, instantiate, and troubleshoot TRL stable post-training trainers: SFT, DPO, GRPO, Reward, and RLOO.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- core-training
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Choose, instantiate, and troubleshoot TRL stable post-training trainers: SFT, DPO, GRPO, Reward, and RLOO.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/trainer-api-reference.md9.2 KB
- 参考文档references/troubleshooting.md4.5 KB
- 参考文档references/workflows.md8.0 KB
- 脚本scripts/inspect_trainer_config.py5.2 KB
03 SKILL.md 正文
Core TRL Training
Use this sub-skill when the task is to select or wire up stable TRL post-training trainers, build trainer objects, set the main config/model arguments, attach reward functions, or debug trainer-specific setup errors.
Route by Training Objective
- SFTTrainer: supervised fine-tuning on text, prompt-completion, or conversational examples; start here for instruction tuning and formatting/loss masking questions.
- DPOTrainer: preference optimization from paired
chosen/rejectedanswers; use after an SFT-style model exists and no online generation rewards are needed. - RewardTrainer: train a sequence-classification reward model from
chosen/rejectedcomparisons; use before RLHF-style training when rewards should come from a model. - GRPOTrainer: online RL from generated completions scored by one or more reward functions or reward models; use for grouped generations and reward shaping.
- RLOOTrainer: online RL with leave-one-out baselines; use when each prompt should produce multiple completions and rewards are compared within that prompt group.
Quick References
- Trainer signatures and default highlights: references/trainer-api-reference.md
- End-to-end trainer skeletons: references/workflows.md
- Setup and runtime failures: references/troubleshooting.md
- Local signature helper: scripts/inspect_trainer_config.py
Boundaries
- Route detailed dataset conversion, chat-template normalization, and reward-function data plumbing to
../data-and-rewards/. - Route CLI command construction, YAML recipes, and launcher flags to
../cli-and-configs/. - Route vLLM, DeepSpeed, FSDP, PEFT scaling, quantization, and memory-distributed execution to
../scaling-and-backends/. - Route experimental trainers, tools/environments, and non-stable APIs to
../experimental-and-environments/.
First Checks Before Instantiating
- Pick the trainer from the objective and dataset shape, then verify the required columns in references/trainer-api-reference.md.
- Create the matching config class rather than passing raw
TrainingArguments, except for simple SFT cases whereSFTTrainercan upgradeTrainingArguments. - Pass a model id string for automatic loading, or pass an already loaded model and keep
model_init_kwargsout of the config. - Keep
trust_remote_code=Falseunless the target model explicitly needs custom code. - Add
peft_configonly when PEFT is installed and the task actually needs adapter training; route scaling details to../scaling-and-backends/.
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).