posttrainer

General-purpose serverless fine-tuning and evaluation for language models.

posttrainer is a RunPod Hub worker that gives you LoRA/QLoRA supervised fine-tuning, lm-eval benchmarking, and LoRA adapter merging — against any HuggingFace model — via a single API call. No infrastructure setup, no pip installs, no waiting for CUDA to compile.

What it does

Task	What happens
train	QLoRA SFT on any HuggingFace dataset via Unsloth + TRL. Pushes merged weights to your HF repo.
eval	Runs lm-eval benchmarks (any task: MMLU, MedQA, HellaSwag, etc.) against any model.
merge	Merges a LoRA adapter into its base model and pushes merged 16-bit weights to HF Hub.

Stack

• Unsloth — 2–5× faster fine-tuning, 70% less VRAM
• TRL — SFT, DPO, GRPO trainers
• PEFT — LoRA, QLoRA, adapter management
• lm-evaluation-harness — 60+ standardised benchmarks
• vLLM — fast inference serving
• Flash Attention 2 — memory-efficient attention
• bitsandbytes — 4-bit and 8-bit quantisation
• Weights & Biases — optional experiment tracking

Usage

Fine-tune a model (QLoRA SFT)

{
  "input": {
    "task": "train",
    "model_id": "unsloth/gemma-4-E4B-it",
    "dataset_id": "FreedomIntelligence/medical-o1-reasoning-SFT",
    "dataset_config": "en",
    "text_field": "text",
    "output_repo": "yourname/my-fine-tuned-model",
    "hf_token": "hf_...",
    "epochs": 1,
    "learning_rate": "2e-4",
    "lora_r": 64,
    "lora_alpha": 64,
    "max_seq_length": 8192,
    "load_in_4bit": true,
    "merge_and_push": true,
    "run_name": "my-run-001",
    "wandb_api_key": "optional"
  }
}

Response:

{
  "status": "completed",
  "output_repo": "yourname/my-fine-tuned-model",
  "train_loss": 1.432,
  "epochs": 1,
  "total_steps": 312
}

Evaluate a model

{
  "input": {
    "task": "eval",
    "model_id": "yourname/my-fine-tuned-model",
    "tasks": ["medqa_4options", "mmlu_clinical_knowledge", "hellaswag"],
    "num_fewshot": 0,
    "limit": 500,
    "batch_size": 4,
    "hf_token": "hf_...",
    "run_name": "eval-001"
  }
}

Response:

{
  "status": "completed",
  "model": "yourname/my-fine-tuned-model",
  "metrics": {
    "medqa_4options/acc,none": 0.6124,
    "mmlu_clinical_knowledge/acc,none": 0.7132,
    "hellaswag/acc_norm,none": 0.8241
  }
}

Merge a LoRA adapter

{
  "input": {
    "task": "merge",
    "base_model_id": "unsloth/gemma-4-E4B-it",
    "adapter_repo": "yourname/my-lora-adapter",
    "output_repo": "yourname/my-merged-model",
    "hf_token": "hf_...",
    "lora_r": 64,
    "lora_alpha": 64
  }
}

GPU recommendations

Model size	Recommended GPU	Notes
Up to 4B	RTX 4090 (24 GB)	QLoRA fits comfortably
7B–8B	RTX A6000 / A100 40GB	QLoRA with seq ≤ 4096
13B–14B	A100 80GB / H100	QLoRA with seq ≤ 4096
30B+	H100 / H200	Multi-GPU or strong QLoRA

Environment variables

All parameters can be passed either in the job input JSON or as environment variables (the Hub UI exposes them with the names below).

Variable	Default	Description
TASK	train	train, eval, or merge
MODEL_ID	—	HuggingFace model ID
HF_TOKEN	—	HuggingFace API token
DATASET_ID	—	HuggingFace dataset ID (train only)
OUTPUT_REPO	—	HF repo to push output to
EPOCHS	1	Training epochs
LORA_R	64	LoRA rank
LORA_ALPHA	64	LoRA alpha
LEARNING_RATE	2e-4	Learning rate
MAX_SEQ_LENGTH	8192	Maximum sequence length
LOAD_IN_4BIT	true	Enable QLoRA (4-bit)
MERGE_AND_PUSH	true	Merge adapter before pushing
EVAL_TASKS	hellaswag,...	Comma-separated lm-eval tasks
WANDB_API_KEY	—	Optional W&B logging

License

Apache 2.0