What the agent does
gensim 4.0.0, released in March 2021, renamed the word2vec constructor's size to vector_size and the model's iter to epochs, and removed the vocab dictionary from KeyedVectors. On gensim 4.4.0 the three failures read TypeError: Word2Vec.__init__() got an unexpected keyword argument 'size', AttributeError: 'Word2Vec' object has no attribute 'iter', and AttributeError: The vocab attribute was removed from KeyedVector in Gensim 4.0.0. Use KeyedVector's .key_to_index dict, .index_to_key list, and methods .get_vecattr(key, attr) and .set_vecattr(key, attr, new_val) instead. A training script written against 3.x hits them in that order, one per run, so the first traceback understates the work.
ChenRocks/fast_abs_rl, the code for an ACL 2018 summarization paper, is a clear case. Its requirements.txt is five lines, pytorch=0.4.0, gensim, tensorboardX, cytoolz and pyrouge; the first is conda syntax, which pip 26.3 refuses with Invalid requirement: 'pytorch=0.4.0' and the hint = is not a valid operator. Did you mean == ?, so a reader installs the packages by hand as the README suggests, and gensim resolves to 4.4.0. Step one of the README's training steps is python train_word2vec.py --path=[path/to/word2vec]. That script builds gensim.models.Word2Vec(size=args.dim, min_count=5, workers=16, sg=1), trains with model.train(sentences, total_examples=model.corpus_count, epochs=model.iter), and names its output files with len(model.wv.vocab)//1000. The embedding loader in utils.py, make_embedding, later opens the saved model with gensim.models.Word2Vec.load(w2v_file).wv and indexes it with w2v[word] and word in w2v, which still work on 4.x.
The agent reproduces on a clean environment first. train_word2vec.py imports count_data from utils.py, which imports torch, so it installs a CPU build of torch alongside gensim and cytoolz, points DATA at a directory with a few small JSON articles in the format the Sentences class reads, and runs the script: it fails at the constructor with the size error. It then applies the three renames, vector_size=args.dim, epochs=model.epochs, and len(model.wv.key_to_index), each the documented 4.0 equivalent, and runs again until the script saves both the .bin model and the .w2v text file.
Then it checks the consumer. make_embedding is run against the file the script just wrote, through the same Word2Vec.load(...).wv path train_abstractor.py uses when given --w2v, to confirm that lookups by word and the in test behave on a 4.x KeyedVectors. The agent does not pin gensim below 4 and adds no hasattr branches: 3.8.3 is from May 2020, and the renamed arguments are the only API this script uses.
Start it with one API call
POST https://api.phoenix.so/o/{org}/sessions with this body:
{
"agent": "claude",
"input": "This repository's word2vec training script fails with \"TypeError: __init__() got an unexpected keyword argument 'size'\" after gensim was installed fresh. Read the dependency files and the README's install instructions, report how gensim is pinned and whether the requirements file is installable with pip as written, then create a clean virtual environment on a current Python, install what the script needs (a CPU build of torch if the script imports it), and report the gensim version that was resolved. Build a small input in the format the script reads, run it, and save the traceback. Search every .py and .ipynb file for the gensim 3 API: size=, iter, wv.vocab, .vocab, index2word, and most_similar on the model instead of wv, and list each use with its file and line. Replace them with the gensim 4 equivalents (vector_size, epochs, key_to_index, index_to_key, wv.most_similar). Do not pin gensim below 4 and do not add try/except or hasattr branches. Run the script again and keep fixing removed gensim APIs until it completes and writes its output files. Then load the saved model through the repository's own loading code and confirm word lookup works. Open a pull request with the list of changed lines, the resolved versions, and the commands that now succeed.",
"resources": [
{
"url": "https://github.com/your-org/your-app"
}
]
}- Get an API key. In the console, open Settings, then API keys, and create a key for your org. Send it in the Authorization header as a Bearer token.
- Send the request. POST the body in request.json to https://api.phoenix.so/o/{org}/sessions, with your repository as the resources URL. The run starts in its own cloud sandbox.
- Read what came back. Follow the run on its live stream or wait for the run.completed webhook, then read the result and its trajectory.
What comes back
A pull request with three renamed tokens in the training script and nothing else, plus a description that carries the proof: the gensim version a clean install resolved, the traceback before the change, every gensim 3 name found in the repository with its file and line, and the python train_word2vec.py --path run that now writes word2vec.128d.0k.bin and its .w2v twin from the small example corpus.
What a reviewer checks is that the renames are exact equivalents and that nothing downstream reads the removed attribute. vector_size, epochs and key_to_index are the names gensim's own error message points to, and the description shows the saved model opened through utils.make_embedding and queried by word. If the repository has notebooks or evaluation scripts that call most_similar on the model object rather than on wv, they are listed as further changes, each its own commit.
A reviewer also sees what was not done: pytorch=0.4.0 is still conda syntax in requirements.txt, and the pull request says so rather than silently rewriting the dependency file, because making it pip-installable is a decision about which torch to name.
Run it on every Monday morning
A repository whose requirements are unpinned breaks on an install day, not on a commit day. Give the Project an automation with a schedule trigger, for example the cron expression 0 7 * * 1 in your timezone, and a prompt that tells the agent to install the dependencies fresh, run each training entry point on a small example input, and open a pull request only when one fails on an API a new release removed. On a week when nothing broke, the run ends with a one-line answer and no pull request.
For a lab's set of paper repositories, start the request above once per repository from a script with an API key, and read each session's pr_url when its run.completed webhook arrives. Each run installs its own torch and gensim in its own sandbox, and runs past the key's max_concurrent limit are accepted and wait as queued.
The use case this job follows, with its automation and trigger: Fix a bug from an alert.
Why run it on Rush?
Three renames take a minute. What takes an afternoon on a laptop is the rest: a clean environment with a CPU torch, an example corpus in the script's format, a run that reaches the save, and the saved model loaded back through the repository's own code. An agent in a sandbox does all of it and leaves every command and its output in the session, so the pull request arrives with the evidence rather than a promise.
It also starts where a new reader starts. A machine that has carried gensim 3.8 since 2020 never shows the error; a sandbox installs what pip install gensim gives today, which is the only environment in which the script fails, and the one in which it has to work.
Frequently asked questions
Why do I get a different error after fixing size?
Because gensim 4 changed three things this script touches, and the script meets them one at a time: the size argument at construction, model.iter at training, and wv.vocab when it names the output file. The agent runs until the script finishes rather than stopping at the first traceback, and lists all three.
Is vector_size the same as size?
Yes, it is a rename of the embedding dimension with the same default. epochs replaces iter the same way, and model.epochs carries the value that model.iter used to, so epochs=model.epochs trains for the same number of passes.
What replaces len(model.wv.vocab)?
len(model.wv.key_to_index), a dict from word to row index that gensim 4 keeps in place of the old vocab dict of Vocab objects. Counts, membership tests and lookups by word all work on it, and index_to_key is the list in the other direction.
Can I pin gensim to 3.8.3 instead?
3.8.3 is from May 2020 and the renames are the only gensim 3 API this script uses, so a pin postpones a three-line change and keeps a five-year-old package next to a current torch. The request above tells the agent not to pin.
Will models saved with gensim 3 still load?
The agent does not test that, because the repository ships no saved word2vec. It trains a fresh one with gensim 4, saves it, and loads it through the repository's own make_embedding; a pretrained 3.x file of your own is worth a follow-up run that loads it the same way.
More gensim jobs
More jobs like this: Fix a bug from an alert
- Fix "RuntimeError: Working outside of application context" from db.create_all() in Flask-SQLAlchemy 3 (Flask-SQLAlchemy)
- Fix "AttributeError: module 'lib' has no attribute 'X509_V_FLAG_CB_ISSUER_CHECK'" From pyOpenSSL (pyOpenSSL)
- Fix "XLRDError: Excel xlsx file; not supported" in a Python Project (xlrd)
- Fix "load() missing 1 required positional argument: 'Loader'" in PyYAML (PyYAML)