Train DeepSeek-V4.1-Flash-1017 with a batch Job

This page runs a managed LoRA fine-tune of deepseek-ai/DeepSeek-V4.1-Flash-1017 with the el CLI. It is submitted exactly like the GLM and Qwen batch Jobs: one JSON request, one Job, one checkpoint. It allocates paid GPUs and is preview, enabled per project: ask EnvLoop support to enable it for your workspace, otherwise submission returns 403 BATCH_POLICY_NOT_APPROVED. See Models and modalities for the limits.

The trainer is NVIDIA NeMo AutoModel; EnvLoop picks and releases the machine.

What to expect

Input	Text conversations only. Images are rejected with 422 INVALID_TRAIN_JOB_SPEC.
Method	LoRA, rank 8, alpha 32 on the attention projections and the routed experts. Not configurable.
Duration	About 45 to 50 minutes for a small dataset, most of it loading the 1.34 TB base weights. Training itself is short: about 1.6 s per step at 512 tokens.
Result	One checkpoint with a LoRA adapter of about 5.5 GB.
Batch size	The Job trains on 8 GPUs in parallel, so one step consumes per_device_train_batch_size × gradient_accumulation_steps × 8 records. 400 records at the defaults make one pass in 50 steps.
Sequence length	max_length defaults to 8192 and longer examples are truncated. Sequences of about 8,200 tokens at per_device_train_batch_size 1 are the validated envelope; larger batches or longer sequences are untested and may run out of GPU memory.
Request size	The complete JSON request must fit 6 MiB.
Capacity	Batch Jobs are admitted one at a time. While another batch Job is active, submission returns 429 TRAIN_JOB_CAPACITY_EXCEEDED; retry later with the same request.
Price	GPU time is billed per second from the moment the machine starts, loading time included. The rate for this GPU is not yet on the Pricing page; read what a run cost with el billing usage list once it has finished.

Prepare the request

Complete installation and authentication, with a project Key that can write training Jobs and download checkpoints and enough balance for the reservation. Update the CLI first (python -m pip install --upgrade envloop-cli) so el job submit --type train is available.

Store your examples as JSONL, one {"messages": [...]} object per line, ending with a non-empty assistant answer. Rows may include system and user messages. For a first test, eight or more rows of made-up facts are enough:

{"messages": [{"role": "user", "content": "How many shards does Orbit Zephora use by default?"}, {"role": "assistant", "content": "Orbit Zephora uses 11 shards by default. -- EnvLoop Orbit Desk"}]}
{"messages": [{"role": "user", "content": "Which TCP port does the Orbit Ilyrin service listen on?"}, {"role": "assistant", "content": "Orbit Ilyrin listens on TCP port 27628. -- EnvLoop Orbit Desk"}]}

Put your rows in train.jsonl and build req.json. Use a new request_id for each new paid run; resubmitting the same file and ID returns the existing Job.

python - <<'PYCODE'
import json
from pathlib import Path

records = [json.loads(line) for line in Path("train.jsonl").read_text().splitlines() if line.strip()]
request = {
    "request_id": "dsv41-orbit-001",
    "base_model": "deepseek-ai/DeepSeek-V4.1-Flash-1017",
    "max_steps": 10,
    "max_length": 512,
    "per_device_train_batch_size": 1,
    "gradient_accumulation_steps": 1,
    "seed": 42,
    "records": records,
}
Path("req.json").write_text(json.dumps(request) + "\n")
PYCODE

Submit and wait

--wait stops waiting after 600 seconds unless you give a limit. Pass at least 3 hours for this model. If the wait ends first, the Job keeps running; read it again with el job get.

el job submit --type train --file req.json --wait --timeout-sec 10800 > result.json

result.json holds the Job; copy its job_id. A passing Job has status completed. To submit without blocking, omit --wait and poll:

el job submit --type train --file req.json > submit.json
el job get YOUR_JOB_ID
el job log YOUR_JOB_ID --source all --follow

Typical phases in the logs: weight download (about 35 minutes), model load, kernel compilation (about 4 minutes), then the training steps. job log --follow ends at the end of capture; Ctrl-C only stops the reader. Stop a Job you no longer want with el job stop YOUR_JOB_ID (billing ends when the machine is released).

Get the result

el checkpoint list --job-id YOUR_JOB_ID --json > checkpoints.json
el checkpoint download YOUR_CHECKPOINT_PATH ./dsv41-adapter --job-id YOUR_JOB_ID

Use the full envloop:// path from checkpoints.json, and a new directory. The download is about 5.5 GB; it verifies every file's size and SHA-256 and can be repeated into the same directory after an interruption. The files are:

File	Content
adapter_model.safetensors	LoRA weights (about 5.5 GB)
adapter_config.json	PEFT adapter configuration
trainer_state.json	global_step and one training loss per step
train_results.json, metadata.json	Run summary; base model, rank and alpha

A healthy run has global_step equal to your max_steps and finite losses that fall over the steps. A falling loss shows the data was fitted, not that answers improved: evaluate on held-out questions. See Checkpoints for listing, retention and deletion.

Limitations

• Text only; no images, audio or tool-call messages.
• LoRA rank 8 only. Full fine-tuning and continued pre-training are not offered.
• Deploying the adapter on an EnvLoop inference endpoint is not available yet: there is no serving profile for this model. Download the adapter and serve it with your own tooling.
• No interactive-SDK training for this model, and no validation split, so there is no validation loss.
• Preview: availability, admission and the GPU configuration may change; check Models and modalities.