BlocksTrainerUsage
Configuration Guide
Trainer config controls data selection, conversion, model choice, training
hyperparameters, infrastructure, and experiment logging. Most runs only need a
few runtime_info.input fields. Leave meta_info, repository pins,
environment paths, and output declarations alone unless you are changing the
block contract itself.
Common fields
| Area | Edit when | Main fields |
|---|---|---|
| Data source | You train from Hugging Face LF data, local LF data, or raw Harbor jobs. | source.type, hf_hub_url, lf_path, job_dir |
| Model | You switch the base model or local checkpoint. | model.model_name_or_path, trust_remote_code |
| Training scale | You tune context length, samples, batch size, learning rate, or epochs. | training.cutoff_len, max_samples, learning_rate, num_train_epochs |
| Infrastructure | You move between one GPU and a multi-GPU node. | infrastructure.n_gpus_per_node, training.deepspeed |
| Remote machine | You want training to run on a GPU server instead of the current host. | meta_info.resources.ip, user, pwd, directory |
| Experiment tracking | You want W&B or Hugging Face access. | experiment.*, credentials.* |
An example reference
Start with the released LF dataset and the default Qwen3.5 base model. This is the simplest path because Trainer can skip Harbor trajectory conversion and go straight to LLaMA-Factory SFT.
runtime_info: input: source: type: hf_lf hf_hub_url: Lego-X/samples_for_llama_factory_sft hf_file_name: jierun_glm52_openswe_mirror_score4_proto_oh_sdk_512_for_qwen3_5.json hf_split: train conversion: max_instances: 0 data_name: glm52_openswe_mirror_score4_proto_oh_sdk_512_qwen3_5 model: model_name_or_path: Qwen/Qwen3.5-35B-A3B-Base trust_remote_code: true training: stage: sft finetuning_type: full template: qwen3_5 cutoff_len: 131072 max_samples: 1000000 output_dir: qwen3_5_35b_a3b_base_glm52_openswe_mirror_score4_proto_oh_sdk_512_gbs64pbs1acc8_lr5e-5_epo3 per_device_train_batch_size: 1 gradient_accumulation_steps: 8 learning_rate: 5.0e-5 num_train_epochs: 3.0 bf16: true flash_attn: fa2 infrastructure: n_gpus_per_node: 8Use /trainer:check after editing. It checks data paths, model settings,
conversion dependencies, GPU configuration, credentials, and training output
paths before /trainer:run starts a costly job.
Configuration breakdown
Remote execution
Trainer runs locally by default, but training often belongs on a GPU server.
Configure meta_info.resources when the block skill should log into a remote
machine and run the Trainer workflow there.
meta_info: resources: ip: 203.0.113.20 # local = current machine; otherwise SSH target user: ubuntu # add this field if your remote setup needs it pwd: null # prefer SSH keys or your private secret channel directory: /home/ubuntu/LegoFlow/blocks/trainer public_storage: nullThe remote directory should point at the Trainer block checkout on the GPU
server. Keep runtime_info.input.infrastructure.n_gpus_per_node aligned with
that server, not with your laptop.
Data source
Trainer can read data in three modes. Choose exactly one source type for the run.
- From a Hugging Face LF dataset:
runtime_info: input: source: type: hf_lf hf_hub_url: Lego-X/samples_for_llama_factory_sft hf_file_name: jierun_glm52_openswe_mirror_score4_proto_oh_sdk_512_for_qwen3_5.json hf_split: train- From a local LF file produced by Tracer:
runtime_info: input: source: type: local_lf lf_path: ../tracer/artifacts/sft_data/<job>/lf.json- From a raw Harbor job directory:
runtime_info: input: source: type: harbor_job scaffold: openhands-sdk job_dir: ../tracer/artifacts/jobs/<job> conversion: max_instances: 0 exclude_repos_file: scripts/excluded_repos.txt data_name: my_converted_datasetUse local_lf when Tracer has already converted trajectories. Use harbor_job
only when Trainer should run conversion itself.
Model and Template
The model and template must agree. For Qwen3.5, keep the current
qwen3_5 template and long context settings.
runtime_info: input: model: model_name_or_path: Qwen/Qwen3.5-35B-A3B-Base trust_remote_code: true training: template: qwen3_5 cutoff_len: 131072 rope_scaling: nullSmoke run
Use a small sample cap and a separate output directory when checking a new data source or training node.
runtime_info: input: training: max_samples: 32 output_dir: smoke-qwen3_5-sft per_device_train_batch_size: 1 gradient_accumulation_steps: 1 num_train_epochs: 1.0 save_steps: 20 infrastructure: n_gpus_per_node: 1Production training
For production, keep the output directory unique and match the GPU count to the node.
runtime_info: input: training: max_samples: 1000000 per_device_train_batch_size: 1 gradient_accumulation_steps: 8 learning_rate: 5.0e-5 num_train_epochs: 3.0 save_steps: 200 bf16: true flash_attn: fa2 infrastructure: n_gpus_per_node: 8Handoff with neighboring blocks
Trainer usually receives converted LF data from Tracer and produces a checkpoint that Evaluator can serve and benchmark.
Receive Tracer data
The most common handoff is local_lf: Tracer writes
artifacts/sft_data/<job>/lf.json, and Trainer reads that file directly.
runtime_info: input: source: type: local_lf lf_path: ../tracer/artifacts/sft_data/<job>/lf.jsonIf you want Trainer to consume raw Harbor jobs instead, keep the dependency
contract and set source.type: harbor_job.
meta_info: dependencies: from: source.job_dir: from: tracer.output.raw_trajectories_dir when: {source.type: harbor_job}runtime_info: input: source: type: harbor_job scaffold: openhands-sdk job_dir: ../tracer/artifacts/jobs/<job>Produce Evaluator checkpoints
After training, use the final checkpoint path as Evaluator's local model input.
runtime_info: input: local_model_serving: model_path: ../trainer/artifacts/<training-output-dir> model_name: Qwen3.5-35B-A3BField reference
| Field | Required? | Default | When to change |
|---|---|---|---|
source.type | Yes | hf_lf | Choose hf_lf, local_lf, or harbor_job for the current run. |
source.hf_hub_url | For hf_lf | released dataset | Change when training from another Hub dataset. |
source.hf_file_name | For file-based hf_lf | released JSON | Change when the dataset repo contains multiple LF files. |
source.lf_path | For local_lf | Tracer LF path | Point to a real lf.json from Tracer. |
source.job_dir | For harbor_job | Tracer jobs path | Point to a real Harbor job directory. |
source.scaffold | For harbor_job | openhands-sdk | Must match the scaffold that produced the trajectories. |
conversion.max_instances | No | 0 | Use a positive number for subsampling; 0 means all samples. |
conversion.data_name | Yes | released dataset name | Change for each converted dataset profile. |
model.model_name_or_path | Yes | Qwen/Qwen3.5-35B-A3B-Base | Change the base model or local checkpoint. |
training.template | Yes | qwen3_5 | Must match the model family. |
training.cutoff_len | Yes | 131072 | Lower for memory checks; keep long for long trajectories. |
training.output_dir | Yes | run name | Change for every important run to avoid overwriting outputs. |
training.learning_rate | Yes | 5.0e-5 | Tune for new data/model settings. |
infrastructure.n_gpus_per_node | Yes | 8 | Must match the training node. |
credentials.* | No | "" | Prefer $HF_TOKEN and $WANDB_API_KEY; never commit secrets. |
meta_info.resources.ip | No | local | Set a host/IP when Trainer should run on a remote GPU server. |
meta_info.resources.user | For remote | add if needed | SSH user for the remote server. |
meta_info.resources.directory | For remote | null | Trainer block working directory on the remote server. |
meta_info.resources.public_storage | No | null | Optional public artifact location for dashboards or sharing. |
Complete config.yaml
Use this as the complete reference after you understand the few fields above.
blocks/trainer/config.yaml
meta_info: name: trainer label: "Trainer" description: Trainer block — SFT training pipeline for SWE-bench coding models. Converts raw agent trajectories from multiple scaffolds into LLaMA-Factory sharegpt format, then trains models via LLaMA-Factory + DeepSpeed (ZeRO-3). parent: legoflow blocks: {} dependencies: from: source.job_dir: from: tracer.output.raw_trajectories_dir when: {source.type: harbor_job} to: {} # nothing currently declares a dependency on trainer's outputs resources: ip: local pwd: null directory: null public_storage: null repositories: llama_factory: url: https://github.com/LegoX/LLaMA-Factory.git branch: main commit: 6e23fa19f485d86d0e5666ce3a9fdfdebb43c050 path: repos/LLaMA-Factory readonly: true swe_data_process: url: https://github.com/LegoX/swe_data_process.git branch: main commit: dda92c581a0c0eaba74bb469d92ddda767e5f1b3 path: repos/swe_data_process readonly: true environment: sft_uv: artifacts/env/lf python_version: "3.12" packages: - repos/LLaMA-Factory - repos/swe_data_processruntime_info: input: source: type: hf_lf # harbor_job | hf_lf | local_lf scaffold: openhands-sdk # harbor_job only: openhands-sdk | claude-code | open-code | terminus2 job_dir: ../tracer/artifacts/jobs/<job> # harbor_job only: replace <job> with a real Tracer job hf_hub_url: "Lego-X/samples_for_llama_factory_sft" hf_file_name: "jierun_glm52_openswe_mirror_score4_proto_oh_sdk_512_for_qwen3_5.json" # empty = full dataset config hf_subset: "" # ignored when hf_file_name is set hf_split: "train" lf_path: ../tracer/artifacts/sft_data/<job>/lf.json # local_lf only: replace <job> with a converted Tracer job conversion: # harbor_job only; for hf_lf/local_lf just data_name + max_instances apply max_instances: 0 # 0 = all samples; >0 = random subsample via num_samples exclude_repos_file: scripts/excluded_repos.txt data_name: "glm52_openswe_mirror_score4_proto_oh_sdk_512_qwen3_5" dataset: name: "" # auto-derived from conversion.data_name model: model_name_or_path: Qwen/Qwen3.5-35B-A3B-Base trust_remote_code: true training: stage: sft finetuning_type: full deepspeed: scripts/deepspeed/ds_z3_config.json template: qwen3_5 cutoff_len: 131072 rope_scaling: null # Qwen3.5 supports 128K natively max_samples: 1000000 preprocessing_num_workers: 16 dataloader_num_workers: 4 output_dir: "qwen3_5_35b_a3b_base_glm52_openswe_mirror_score4_proto_oh_sdk_512_gbs64pbs1acc8_lr5e-5_epo3" logging_steps: 1 save_strategy: steps save_steps: 200 overwrite_output_dir: false # explicit env override required to replace an existing run save_only_model: true resume_from_checkpoint: null per_device_train_batch_size: 1 gradient_accumulation_steps: 8 learning_rate: 5.0e-5 weight_decay: 0.01 max_grad_norm: 1.0 num_train_epochs: 3.0 lr_scheduler_type: cosine warmup_ratio: 0.1 bf16: true ddp_timeout: 180000000 enable_liger_kernel: true use_unsloth_gc: true flash_attn: fa2 infrastructure: n_gpus_per_node: 8 experiment: run_name: "" # "" = auto-derived wandb_mode: offline wandb_run_id: "" # "" = auto-assigned credentials: wandb_api_key: "" # via $WANDB_API_KEY hf_token: "" # via $HF_TOKEN; needed only for private hf_lf datasets output: checkpoint_path: description: Path to the latest trained model checkpoint value: null artifacts: train_results: null train_loss_plot: null training_log: null training_curves: description: WandB run id with loss, learning rate, and token metrics value: null training_metrics: description: Final training metrics from LLaMA-Factory value: null