Models and training methods

Base models available in alallo

alallo fine-tunes open, publicly available instruction-tuned models. The alallo model catalogue currently contains:

Model in the dashboard	Hugging Face ID	Size (parameters)
Qwen2.5 0.5B Instruct	Qwen/Qwen2.5-0.5B-Instruct	0.5 billion
Qwen2.5 1.5B Instruct	Qwen/Qwen2.5-1.5B-Instruct	1.5 billion
Qwen2.5 3B Instruct	Qwen/Qwen2.5-3B-Instruct	3.1 billion
Phi-3.5 mini Instruct (3.8B)	microsoft/Phi-3.5-mini-instruct	3.8 billion
Qwen2.5 7B Instruct	Qwen/Qwen2.5-7B-Instruct	7.6 billion
Qwen2.5 14B Instruct	Qwen/Qwen2.5-14B-Instruct	14.8 billion

"B" means billion parameters. A larger model can learn more complex behaviour but takes longer to train and uses more alallo tokens per training token. [TBD: Qwen2.5 32B Instruct is in the catalogue but is not offered yet; confirm whether and when larger models will be offered.] [TBD: whether customers can request other models.]

Why an alallo model or method shows "Not available right now"

The alallo dashboard only offers combinations of model and training method that alallo can run at the moment for the data residency you chose. The Base model list only contains models with at least one available method. A method card or a precision option that cannot run for the selected model shows "Not available right now". When full precision (16-bit) is not available for a model but 4-bit is, the Supervised fine-tuning card says "This model can't run at full precision right now. Choose a lower precision." Availability can differ between the EU, US and Anywhere data residency options, and it can change over time.

Training methods in alallo, in plain language

A training method (called "Algorithm" in the dashboard) decides how the base model learns from your data. alallo offers four methods:

Method ID	Dashboard name	Data it needs
sft-lora	Supervised fine-tuning, Precision "Full precision (16-bit)"	Text examples
sft-qlora	Supervised fine-tuning, Precision "4-bit"	Text examples
sft-full	Supervised fine-tuning (full weights)	Text examples
dpo-lora	Direct preference optimisation (LoRA)	Prompt, chosen answer, rejected answer

Supervised fine-tuning (SFT) in alallo

Supervised fine-tuning (SFT) teaches a model by showing it example texts, for example an instruction followed by the ideal answer. The model learns to produce text like your examples. Use supervised fine-tuning when you have examples of the output you want. alallo offers three kinds of supervised fine-tuning: LoRA, QLoRA and full weights.

LoRA fine-tuning in alallo (Supervised fine-tuning, full precision 16-bit)

LoRA (Low-Rank Adaptation) leaves the base model's original weights untouched and trains a small set of extra weights, called an adapter, on top of them. alallo describes LoRA as: "Trains small low-rank adapters on instruction or text data. Cheap, fast, small download." The result of a LoRA job is a small adapter file that you use together with the original base model. LoRA is the recommended starting point for most alallo customers. In the dashboard, LoRA is the "Supervised fine-tuning" card with Precision set to "Full precision (16-bit)".

QLoRA fine-tuning in alallo (Supervised fine-tuning, 4-bit)

QLoRA is LoRA on a compressed (4-bit quantised) copy of the base model. alallo describes QLoRA as: "LoRA on a 4-bit quantised base model. Trains larger models for less." Compressing the base model lets larger models train when full precision is not available, at a possible small cost in quality compared with 16-bit LoRA. The result of a QLoRA job is a LoRA adapter, like a LoRA job. In the dashboard, QLoRA is the "Supervised fine-tuning" card with Precision set to "4-bit".

Full fine-tuning in alallo (Supervised fine-tuning, full weights)

Full fine-tuning updates every weight of the base model instead of training an adapter. alallo describes full fine-tuning as: "Updates every weight. Highest quality ceiling; costs the most." The result of a full fine-tuning job is a complete model, not an adapter. Full fine-tuning has a lower default learning rate (0.00002) than the LoRA methods. Use full fine-tuning when LoRA does not reach the quality you need.

DPO in alallo (Direct preference optimisation, LoRA)

Direct preference optimisation (DPO) teaches a model which of two answers to prefer. Each training example has a prompt, a "chosen" (better) answer and a "rejected" (worse) answer. alallo describes DPO as: "Learns from chosen vs rejected answers to the same prompt. The frozen base model is the reference, so no second copy is loaded." DPO in alallo trains a LoRA adapter. Use DPO to steer tone, safety or answer quality when you can say which answers are better, rather than writing the perfect answer yourself. DPO is often run on a model that has already been supervised fine-tuned. [TBD: whether a customer can run DPO on top of their own earlier alallo fine-tune.]

Common training settings in alallo

Every alallo training method has these settings. The defaults work for a first job.

Setting	Default	Allowed range	What it does
Max steps	200	1 to 1,000,000	How many training steps the job runs. More steps means more training, a higher price and a longer time.
Learning rate	0.0002 (full weights: 0.00002)	0.0000001 to 1.0	How big each learning update is. Too high can make training unstable; too low makes it learn slowly.
Batch size	4	1 to 512	How many examples are processed together in one pass.
Gradient accumulation	4	1 to 512	How many passes are combined into one training step. Each step learns from batch size × gradient accumulation sequences (16 with the defaults).
Max sequence length	512	16 to 32,768	The length, in tokens, of each training sequence. Longer texts need a longer sequence length.
Checkpoint every N steps	50	1 to 100,000	How often training progress is saved. "Lower means less work is redone if training is interrupted."

How much data an alallo job trains on

The amount of training in an alallo job is set by Max steps, not by the size of your dataset. Each step processes batch size × gradient accumulation × max sequence length tokens (with defaults: 4 × 4 × 512 = 8,192 tokens per step). If your data has fewer tokens than the job processes, alallo goes through your data more than once. If your data has more, the job uses part of it. For supervised fine-tuning, alallo joins your examples into full-length sequences, so no compute is wasted on padding; each example still only "sees" itself during training.

LoRA settings in alallo

The LoRA, QLoRA and DPO methods have these extra settings:

Setting	Default	Allowed range	What it does
LoRA rank	16	1 to 1024	The size of the adapter. A higher rank can learn more but makes a larger adapter.
LoRA alpha	32	1 to 4096	How strongly the adapter changes the model's output. Often set to twice the rank.
LoRA dropout	0.05	0.0 to 0.9	Randomly ignores part of the adapter during training to reduce overfitting.

DPO beta setting in alallo

The DPO method has one extra setting, DPO beta (default 0.1, allowed range 0.001 to 10.0). alallo describes it as: "How far the policy may drift from the reference model." A lower beta lets the fine-tuned model move further from the base model's behaviour; a higher beta keeps it closer.

The "Require bf16" option in alallo

The "Require bf16" checkbox in the Algorithm section is on by default. bf16 and fp16 are two 16-bit number formats used during training. The dashboard explains: "More numerically stable, especially for larger models. Turn off to allow fp16, which can start sooner." Leave "Require bf16" on unless you want the job to start sooner and accept a small risk of less stable training.

Choosing an alallo method: quick guide

• You have examples of good outputs and want the model to imitate them: Supervised fine-tuning, full precision (16-bit).
• The model you want is not available at full precision: Supervised fine-tuning, 4-bit.
• You need the highest possible quality and LoRA was not enough: Supervised fine-tuning (full weights).
• You have pairs of better and worse answers: Direct preference optimisation (LoRA).