LLaMA3.2-3B-Instruct-talkie-distill

LLaMA-3.2-3B-it-talkie

This repository has a distilled version of LLaMA-3.2-3B, trained on my 1M token sized Talkie dataset for two epochs. Just as with my last talkie-distill release the safetensor version of this model will be available as well and the configuration i used as well as more.

The reason for the upscale from the last, Qwen3.5-0.8B model to this one is to see if a larger model handles it better, and it actually seems to. It's able to generalize it's new training data and way of speaking with moden concepts, even if it does get things more inaccurately (due to it's training data).

*NOTE: This is a 'LoRA' finetune.

Here is some examples of that generalization:

INPUT: What is a large language model?

OUTPUT: A large language model is a statistical model for representing
a language by a set of statistical associations between
characters. The model is built on the assumption that
each character can be defined as a probability distribution
over a large number of characters, and that the distribu-
tions for different characters are independent and identi-
cal. The model must, therefore, be a large, one, and
can be regarded as an aggregate of many smaller models,
each corresponding to one of the characters of the language.
The model must be built up gradually, step by step, by
proceeding from the simplest to the most complex, so that
at each stage the total number of characters shall be
increased by a small increment.


INPUT: What is tokenization and back propagation?

OUTPUT: Tokenization and back propagation are the processes by
which the model is fitted to the language data in hand.
Tokenization means the separation of the language into
characters, and back propagation means the process by
which the statistical weights are modified so as to minimise
the difference between the model and the language data.
The process of fitting the model to the language data
consists in adjusting the statistical weights in such a way
that the output of the model shall be most like the actual
language data. The differences between the model and the
data are calculated in terms of some measure of disti-
bution (e.g. the sum of squares of errors of prediction),
and these are made minimum by an iterative procedure
in which the model is gradually changed, and the
difference is recalculated, until no further improvement
is perceptible.

It does confuse our modern knowledge of what LLMs, Tokenization, and backprop are (this is to be expected with this distillation.), but it does manage to essentially merge it's knowledge with the style it was trained on.

Since this model is LLaMA you are beholden to the LLaMA-3.2 Community License Agreement, It is not my choice. I will soon release a alternative Qwen3.5-4B distillation for those who want a more liberating license. As always, Credit is given to the talkie team for creating Talkie, and releasing it to the public. The dataset this model was trained on is publicly available, if you want to run your own distillation.