Which job did that fine-tune finish?

Do not treat a model name as the job a fine-tune finished, and do not ask the model which job trained it, until you can name the job. Tonight, three proofs, in the log you stored. Name the job that fine-tune finished: the job id, the model, and the status. If fine_tuned_model is null, write that, and do not invent a model name to fill the hole. Show a different job whose id, or whose model, is not this one. Show another organization's list that does not hold this job. If any of the three needs the person who uploaded the file and then left, you have not named the job. You have a model string and a guess.

This is one fine-tuning job. It is not the batch a request joined. Which batch did that request join? is a batch id and a custom id. A batch can finish while fine_tuned_model is still null. It is not the model a single request hit. Name the model the request hit is the model on one call. A base model on a job is not proof the job succeeded. It is not the file a message attached. Which file did that message attach? is a file on a message. A training_file id can equal a result file id in a sample, and the two roles are still not one field. It is not the voice a speech request used. Which voice did that speech use? is a voice on one audio request. A voice is not a job id. It is not the setup a customer confirmed, and it is not the hook a repository called. Which setup did the customer confirm? is a setup id. Which hook did that repository call? is a hook id.

None of what follows is a score for the model you are rescuing, and none of it is legal advice. The fields below are the current examples in the OpenAI OpenAPI description of fine-tuning jobs. Several examples reuse one id and do not reuse the model, the seed, or the created time. The figures below that are a 2021 paper about low-rank adaptation. A percentage of someone else's parameters is not the job id in the log you stored.

The first create example does not agree with itself across languages, and the response agrees with only one of them. The curl sends training_file file-BK7bzQj3FfZFXr7DbL6xJwfo. 29 characters. The Python sample sends file-abc123. 11 characters. The JavaScript sample sends file-abc123 and does not send a model. The curl and the Python sample send model gpt-4o-mini. 11 characters. The response id is ftjob-abc123. 12 characters. The response model is gpt-4o-mini-2024-07-18. 22 characters, not 11. The response training_file is file-abc123, which matches Python and does not match the curl. status is queued. 6 characters. fine_tuned_model is null. finished_at is null. seed is null. created_at is 1721764800. organization_id is org-123. 7 characters. object is fine_tuning.job. 15 characters. method type is supervised. 10 characters. n_epochs is the string auto. 4 characters. error is an empty object. The schema allows an error object, null, or an empty object. Empty is what this response stored (OpenAI, openapi.yaml, fine-tuning jobs).

The epochs example reuses the id and the created time and changes the model and the seed. The response model is gpt-4o-mini, the 11-character request name, not the 22-character snapshot. n_epochs is the number 2, not the string auto. seed is 683058546, not null. The same id, on the first response, had a null seed. The schema says n_epochs is either auto or an integer from 1 to 50. 2 is inside that range. 50 is not 2. batch_size is auto here. The schema says batch_size is auto or an integer from 1 to 256. The retrieve example, below, uses batch_size 1, the minimum, not 256.

The DPO example is a different id, and the request languages do not share files. The curl sends training_file and validation_file both as file-abc123. The Python sample sends training_file file-abc and validation_file file-123. Both of those are 8 characters, not 11. The response id is ftjob-abc. 9 characters, not 12. created_at is 1746130590. Subtract 1721764800 and the gap is 24365790 seconds. 86400 times 282 is 24364800. The remainder is 990 seconds, which is 16 minutes and 30 seconds. The DPO sample was created 282 days, 16 minutes, and 30 seconds after the queued sample. organization_id is org-abc. 7 characters, the same length as org-123, a different id. status is queued. beta is 0.1. seed is 1036326793. validation_file is file-123. Do not paste this created time onto ftjob-abc123 because both are fine-tuning jobs.

The reinforcement example reuses ftjob-abc123 and then disagrees with the request. The curl grader input is {{sample.output_text}}. 22 characters. The reference is {{item.label}}. 14 characters. The Python sample swaps them: input is {{item.label}} and reference is {{sample.output_text}}. The response matches the curl, not the Python. The Python call also sends seed 42. The response seed is 950189191. 42 is not 950189191. status is validating_files. 16 characters, not queued. model is o4-mini. 7 characters. The grader name is Example string check grader. 27 characters. operation is eq. reasoning_effort is medium. 6 characters. created_at is again 1721764800, the queued sample's clock, on a status that is not queued.

The list, the retrieve, and the cancel each pick a different fact for the same short id. The list default limit is 20. The example request sends limit=2. The response data holds 1 job. has_more is false. 1 is not 2, and 2 is not 20. That one job is ftjob-abc123, model gpt-4o-mini-2024-07-18, created_at 1721764800, status queued, seed 42, error null. The create response's error was an empty object, and its seed was null. null is not an empty object, and 42 is not null. The retrieve curl asks for ft-AF1WoRqd3aJAHsqc9NY7iL8F. 27 characters. The Python and JavaScript samples, and the response, use ftjob-abc123. 12 is not 27. The response model is davinci-002. 11 characters. created_at is 1692661014. finished_at is 1692661190. The difference is 176 seconds, which is 2 minutes and 56 seconds. status is succeeded. 9 characters. fine_tuned_model is ft:davinci-002:my-org:custom_suffix:7q8mpxmy. 44 characters. trained_tokens is 5768. n_epochs is 4. batch_size is 1. learning_rate_multiplier is 1.0. seed is 0. estimated_finish is 0. 0 is not 1692661190. result_files holds file-abc123, the same 11 characters as training_file. Same string, two fields. Subtract the retrieve created time from 1721764800 and the gap is 29103786 seconds. 86400 times 336 is 29030400. The remainder is 73386 seconds: 20 hours, 23 minutes, and 6 seconds. Same printed id, 336 days, 20 hours, 23 minutes, and 6 seconds apart. Do not paste davinci-002 onto the queued snapshot because the id string matches.

Cancel, checkpoints, and events keep teaching the same lesson. Cancel returns status cancelled. 9 characters, two L characters, which is not the one-L spelling a card network uses on a different object. hyperparameters is null. finished_at is null. validation_file is file-abc123, where the first create response had validation_file null. A cancelled job in this sample has no finished time. The checkpoint list default limit is 10. The sample returns 2, and has_more is true, so 2 is not the whole set. The first checkpoint id is ftckpt_zc4Q7MP6XxulcVzj4MZdwsAB. 31 characters. created_at 1721764867. step_number 2000. full_valid_loss 0.134. full_valid_mean_token_accuracy 0.874. The second checkpoint id is ftckpt_enQCFmOTGj3syEpYVhBRLTSy. 31 characters. created_at 1721764800, equal to the queued job. step_number 1000. loss 0.167. accuracy 0.781. 1721764867 minus 1721764800 is 67 seconds. The later step is the one with the lower loss. 0.167 minus 0.134 is 0.033. 0.874 minus 0.781 is 0.093. The suffix 7q8mpxmy appears on the step-1000 checkpoint name and also on the retrieve example's davinci model. The models around that suffix are not the same. The events example asks for limit 2 and returns 2, with has_more true. Both events have created_at 1721764800. One message says the job successfully completed. The other says a new model was created, ft:gpt-4o-mini:openai::7p4lURel. 31 characters. The queued job at that same timestamp has status queued and fine_tuned_model null. The event is not the job. integrations allows at most 5. The Weights and Biases example sends 1. The request as printed is missing the comma after the name field, so it is not valid JSON. The response drops the name and the tags, sets entity to null, and sets run_id to ftjob-abc123, the job id. A display name in a request is not a field the response kept.

Hu, Shen, Wallis, Allen-Zhu, Li, Wang, Wang, and Chen measured adaptation rank, which is not this job. They use GPT-3 175B as the example, and they say the trainable parameters can be as small as 0.01% of the original. 175 billion times 0.01 percent is 17.5 million. That sentence does not print 17.5 million. They say a rank r of one or two suffices when the full dimension d is 12288. 12288 divided by 2 is 6144. They say training VRAM on that model drops from 1.2 TB to 350 GB. Treating the units the way their next ratio treats gigabytes and megabytes, 1.2 TB is 1200 GB. Two thirds of 1200 is 800, and 1200 minus 800 is 400. The GPT-3 sentence prints 350, not 400. The checkpoint line says the size drops by roughly 10000 times, from 350 GB to 35 MB. 350 divided by 0.035 is 10000. 35 MB times 100 is 3500 MB, which is 3.5 GB, and 350 plus 3.5 is 353.5. The paper prints about 354 GB for one hundred adapted models. 353.5 is not 354. The speed sentence says a 25% speedup. The footnote prints 32.5 tokens per second for full fine-tuning and 43.1 for LoRA, per V100. 32.5 times 1.25 is 40.625. 43.1 is not 40.625. 43.1 minus 32.5 is 10.6. 10.6 divided by 32.5 is 0.326, which is 32.6% to one decimal, not 25%. The paper does not print 40.625. Table 1 is a latency table on GPT-2 medium, averaged over 100 trials. 100 is not the epoch maximum of 50, and it is not job id ftjob-abc123 (Edward Hu and colleagues, LoRA: Low-Rank Adaptation of Large Language Models).

Inflectiv is a platform of domain assistants, distinct from the next assistant. It is not a fine-tuning job ledger. No headcount, duration, or result figure belongs in that sentence. Inflectiv Helios does not supply ftjob-abc123.

When the pilot is real and the log cannot name the job, the missing object is specific. A record you stored, that shows one job id, one model, and one status, or an explicit null fine-tuned model, shows a different job that is not that one, and shows another organization's list without it. An AI pilot failed to reach production engagement is an audit in weeks 1 to 2, then an engagement finished in 8 to 12 weeks, scoped after the audit, unpaid if the agreed metric does not move, with rollout behind a human-review gate. The harness that scores a stored job id against the model the request names, on your examples, is AI evaluation and observability. An AI integration sprint is the shape where the acceptance sentence can require the three proofs on every stored job. If you do not yet know whether the pilot can be saved, a two-week AI readiness assessment hands you the findings either way. Those are the terms. They are not a promise that a 25% footnote will match 43.1 divided by 32.5.

There is a version of this afternoon where a rescue is the wrong purchase. Nothing is in production. No model is being called. The request was not logged, and there is no job id either. A customer has already seen another customer's files. The wrong value was already a write. Or no metric was ever named. A model name you cannot tie to a job is unpleasant. It is not yet a system you can measure. 0.01% is not your job, and 5768 is not your token count.

Before you tell them which job the fine-tune finished, do the three proofs and write the name. Name the id, the model, and the status, or write that the fine-tuned model was null. Show a different job that is not that one. Show another organization's list that does not hold it. Do not paste a snapshot onto a request that named the short model, and do not ask a model which job it was. The binding is the job you can defend.