// Direct Preference Optimization (DPO) fine-tune with TRL `DPOTrainer`. Triggered when the user wants to align a model on preferences / pairwise comparisons / chosen-vs-rejected data, or improve an existing SFT checkpoint with a preference dataset.
Set up training and inference on AMD Ryzen AI Max+ 395 / Strix Halo (gfx1151, RDNA 3.5) with TheRock nightly ROCm wheels. Triggered when the host has gfx1151, when `rocminfo` shows Strix Halo, or when the user mentions Strix Halo / Ryzen AI Max / gfx1151 / 128GB unified memory.
Recreate the ml-intern-kit Python environment on a new machine (laptop, rented GPU box, Docker, fresh checkout). Triggered when the user is on a new host, sees ImportError on a core dep (torch/transformers/trl/peft/accelerate), or wants to install flash-attn / unsloth / bitsandbytes after the fact.
Evaluate a trained or downloaded language model with `lm-eval-harness` standard tasks (arc, hellaswag, gsm8k, mmlu, truthfulqa, ifeval, ...). Triggered when the user wants to benchmark, eval, or compare a model — pre- or post-training.
Audit a Hugging Face dataset before training to confirm splits, columns, format, sample rows, distributions, and duplicates. Triggered before any training/fine-tuning script runs, when a user mentions a new dataset, or when you hit a KeyError / format mismatch in a training job.
Submit a training/inference script to Hugging Face Jobs (`hf jobs run`). Triggered when the user wants to run training in the cloud, scale beyond local hardware, or kick off a multi-hour fine-tune. Enforces pre-flight: hub_model_id required, ≥2h timeout for training, single-job validation before batch.
Run a literature-first crawl before writing ANY ML training/fine-tuning/inference code. Spawns an Explore sub-agent that mines papers, citation graphs, methodology sections, and matched HF datasets to produce a ranked list of training recipes attributed to specific published results. Triggered when the user asks to fine-tune, train, or improve a model, or when the user names a task/benchmark and you need a recipe before coding.
| name | train-dpo |
| description | Direct Preference Optimization (DPO) fine-tune with TRL `DPOTrainer`. Triggered when the user wants to align a model on preferences / pairwise comparisons / chosen-vs-rejected data, or improve an existing SFT checkpoint with a preference dataset. |
Template: scripts/train_dpo.py. Required dataset columns: prompt,
chosen, rejected.
python scripts/inspect_dataset.py <dataset>. Verify
the three required columns exist and aren't empty / duplicated.configs/sft_default.yaml; for DPO add/override:
train:
beta: 0.1 # KL coefficient — typical range 0.05 – 0.5
max_length: 2048
max_prompt_length: 1024
learning_rate: 5.0e-7 # DPO is much more sensitive than SFT
num_train_epochs: 1
--max-steps 20 --max-samples 256 discipline.rewards/margin stays at 0
the dataset is uninformative. If KL blows up, lower the learning rate.accelerate or via scripts/launch_hf_job.py.DPO doubles GPU memory (policy + reference model). Size up one tier from what you'd pick for SFT of the same model.