This model follows roughly follows the Humpback architecture, proposed in the paper Self-Alignment with Instruction Backtranslation by Li et al. An additional improvement, primarily inspired by the paper Better Alignment with Instruction Back-and-Forth Translation by Nguyen et al., is added at the end of the original pipeline.

The original Humpback uses instruction backtranslation on a web corpus to generate input-output pairs (self-augmentation), creating a richer dataset for fine-tuning models without the need for additional manual annotation. For this, the documents from the web corpus are treated as theoretical responses, for which then matching instructions are generated. A copy of the base model, instruction-tuned on a small amount of "gold" instruction-response pairs, then iteratively curates the created dataset, scoring the pairs by quality, and is then finetuned on the resulting subset of all pairs with the highest possible score (self-curation). The pipeline by Nguyen et al. adds a third step called "Rewriting". During this step an already aligned LLM (e.g. LLaMa-2-70B-chat) is employed to rewrite those responses that have passed the filtering at the self-curation step. The rewriting improves the linguistic quality of the responses, due to the nature of web-sourced texts, often containing colloquialisms and stylistic noise. The final model is then finetuned on the rewritten dataset.

This approach inspired me to also add a rewriting step, performed not by an already aligned external LLM, but by the "seed model", that also performs the filtering (self-curation). This approach intends to bring back the idea of "Self-Alignment", since using an external model for rewriting deviates from the "self" aspect. In my pipeline the "self-rewriting" step is performed before self-curation, so that the quality of the pairs is ensured after rewriting, allowing for more candidate pairs to be taken into consideration during filtering. This can be important for leveraging the amount of data used, since some web documents have messy structure and would get filtered out when performing filtering first. The rewriting could potentially restructure the response and thereby increase its quality and chance to be included in the final training data, potentially allowing for a greater, more diverse final training dataset.

This model represents the resulting model after the first iteration of the pipeline, which is trained on a small amount of gold data and a set of generated data rewritten and curated by the "seed model".

This model can be used for instruction-following. It may also be used to, again, rewrite and score the instruction-response pairs generated by the "backward model" for a second iteration of the pipeline.

Varying from the original paper, this model is a fine-tuned version of meta-llama/Llama-3.1-8B. It has been trained using TRL.

The dataset used to train this model is a combination of data sampled from the oasst1 dataset and the synthetic dataset which was mentioned above. The latter has been created by applying self-augmentation, self-rewriting and self-curation on 502k entries from the english subset ("en") of the c4 dataset.

Original paper:

Inspiration:

Cite TRL as: