Nebius Token Factory fine-tuning

# Fix list: Nebius Token Factory fine-tuning

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/nebius-token-factory-fine-tuning, the October 2026 research run, assessed 8 October 2026. Grade D, 47.7 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Nebius Token Factory fine-tuning: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Payments & pricing, 0 out of 100, up to 12.5 more on the total

Why it scored 0: No machine payment protocol (0). No fine-tuning price was found in the docs or the public catalogue JSON, and the product page points to a console price page that is script-drawn and disallowed by robots.txt, which we did not read. Scored as absent for that reason, not as a finding that prices are hidden (0). A bank card is mandatory at onboarding, so the $1 credit for 30 days is not a card-free trial (0). Sign-up, billing and key creation are browser steps (0).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 2. Reliability, 45 out of 100, up to 11 more on the total

Why it scored 45: Scored with the hosted lines. status.nebius.com is a Statuspage for the whole Nebius cloud with a Token Factory component in each of nine regions and none for fine-tuning (15 of 20). From 10 July to 8 October 2026 two major incidents were tagged Token Factory, a network incident of 93 minutes on 27 July and overheating hardware in us-central1 from 19 to 20 August lasting about 21.5 hours, with a post-mortem linked on 1 September. The second was confined to one region, so 5 instead of 0. Rate limits are dynamic, the docs show a 60 requests and 400,000 tokens a minute baseline as an example and send readers to the console for their own, and no fine-tuning job limit was found (8). 429 responses carry `Retry-After` and `x-ratelimit-*` headers, and the guide says to poll at 15 seconds or slower and to recreate a job after a 5xx, with no idempotency key (10). The published SLA document names no Token Factory service level, and the docs say an SLA comes with the Enterprise tier by contract (0). Supervised fine-tuning carries no beta label, but custom model weights for serving are in beta on request (7).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 3. Security & auth, 52 out of 100, up to 8.4 more on the total

Why it scored 52: Bearer API keys created and deleted per project in the console, shown once. No scopes, expiry or rotation API found in the reviewed documentation (20). Project Admins and Members both have full access to the Files and Fine-tuning APIs, the Billing Manager group has no resource access, and there is no read-only role or confirmation step for deletes (5). Responses carry job state and the customer's own files, no third-party content (10). A usage view by project, service and product. No audit log for Token Factory API calls found (3). security.txt on nebius.com is valid until 31 December 2027, and the trust centre lists SOC 2 Type II, ISO 27001, 27701 and 42001 among others without naming Token Factory in scope. No bug bounty found (14). Hugging Face and Weights & Biases tokens travel in the job request body and are marked write-only in the spec.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 4. Agent ergonomics, 51 out of 100, up to 8 more on the total

Why it scored 51: List calls take `limit`, and job objects are compact, with no field selection (12). `limit` and `after` cursors on jobs, events and checkpoints, and a `purpose` filter on files. No filter by status or model found (12). A failed job returns an `error` object with `code`, `message` and `param`, and validation errors name the field. No list of error codes found (10). No idempotency key on job creation. The guide says to recreate a job after a 5xx, and a cancel call exists (5). Only `model` and `training_file` are required and every hyperparameter has a default. There is no Nebius SDK for this API, and the docs use OpenAI's Python and JavaScript clients (12).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 5. Schema & documentation, 68 out of 100, up to 5.2 more on the total

Why it scored 68: A public OpenAPI 3.1 file at api.tokenfactory.nebius.com/openapi.json covers `/v1/fine_tuning/jobs`, events, checkpoints, files, datasets and operations (25). llms.txt, llms-full.txt and a Markdown twin of every docs page (10). Operation descriptions are one line each. The guide explains each hyperparameter with ranges and when to change it (10). Hyperparameters carry minimums, maximums and defaults, status is a six-value enum and two fields are required, but the request allows additional properties and a free-form `extra_body` (11). Python and cURL examples with sample responses. The reference documents only a 422 validation response, and the guide disagrees with the spec on integrations (it lists only `wandb` as supported and omits the required `api_key`) (7). Paths are versioned `/v1` and the spec is date-stamped 20260930, with no public changelog found (5).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 6. Maintenance & community, 61 out of 100, up to 3.4 more on the total

Why it scored 61: The OpenAPI file is stamped 20260930 and the fine-tuning reference pages were last modified on 2 October 2026 per the docs sitemap (30). No changelog exists. Dated evidence in the last 90 days is the August deprecation notice, legal documents of 15 and 23 September and 177 docs pages modified in September and October, which earns half (10). Support is a contact form and a Discord community, and the cookbook repository had 127 commits since 10 July 2026 with merged pull requests. We did not read its issue tracker (8). No SDK of its own. The docs rely on current OpenAI clients (8). The cookbook carries tests and a SECURITY.md that points to a different repository (5).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## 7. Transparency & trust, 79 out of 100, up to 1.8 more on the total

Made of editorial 70, provenance 88.

Why it scored 79: Closed service under the Nebius Services Agreement and Token Factory Supplemental Terms, which name the contracting entity by region. The cookbook is MIT (18). The legal guide, DPA and terms agree that content is not used for training and that fine-tuning data is stored in the EU. Inputs and outputs are kept for speculative decoding unless Zero Data Retention is on, no retention period for datasets or checkpoints was found, and the terms allow deletion of tuned models at three days' notice (22). Dated deprecation notices for June and August 2026 name replacements for serverless models. No written minimum notice period found (10). A Token Factory sub-processor list dated 23 September 2026 gives each entity, role, location and transfer mechanism, with 15 days' notice of changes (20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Terms of service: read, states 7 of the 7 things a reader expects, and has 1 clause that costs points (8 of 10)
- Changelog: not found (0 of 10)

## Deductions

Each comes off the total. A fixed and documented problem counts for less at the next check.

- -2: on 2026-10-08 the product page said a fine-tuned model goes live with one click on 'serverless endpoints, on-demand GPU, or dedicated enterprise clusters', while the docs say deployment is by dedicated endpoints only and custom model weights are in beta on request, and the guide's deployment links return 404 (https://nebius.com/services/token-factory/fine-tuning, https://docs.tokenfactory.nebius.com/post-training/models.md)

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked: the fine-tuning price list. The product page points to https://tokenfactory.nebius.com/organization/prices, which is drawn by script inside the console and disallowed by robots.txt, so Payments scores pricing as absent on our inability to read it.
- unchecked: GitHub stars and the issue tracker of nebius/token-factory-cookbook. The GitHub API refused the one request we made.
- Whether serverless LoRA deployment with per-token billing still exists. An older docs page and the product page describe it, the current models page says dedicated endpoints only, and the deployment guide returns 404.
- Whether the trust centre's SOC 2 and ISO certificates cover Token Factory. The legal guide says so in general terms and the trust centre page does not name the product.
- Whether failed or cancelled fine-tuning jobs are billed, and how many jobs can run at once, were not found in the reviewed documentation.
- The trust centre and product page were read through a summarising fetch. The product page wording on deployment was confirmed against the raw HTML.
- The lead's docs link (`fine-tuning/overview`) still loads but is an older page outside the docs index. The current section is `post-training/overview`.

## Weaknesses

- No fine-tuning price found in the docs or the public catalogue JSON. The price page is a script-drawn console page that robots.txt disallows
- The models page says deployment is by dedicated endpoints only, and custom model weights are in beta and available on request
- A bank card is mandatory at onboarding, so the $1 trial credit (30 days) is not a card-free trial
- Two major incidents tagged Token Factory in 90 days, 93 minutes on 27 July and about 21.5 hours in us-central1 from 19 August 2026
- No public changelog, no idempotency key on job creation and only a 422 response documented in the reference
- The Services Agreement (clause 4.1.10) forbids competitive analysis or benchmarking

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Use the OpenAI client with `base_url` `https://api.tokenfactory.nebius.com/v1/` and `NEBIUS_API_KEY`. Upload JSONL with `purpose=fine-tune`, then create the job.
- Set `hyperparameters.lora` to true for an adapter. The default is false, which runs full fine-tuning.
- Poll `GET /v1/fine_tuning/jobs/{job_id}` no faster than every 15 seconds. There is no idempotency key, so list jobs before recreating one after a timeout.
- Download every ID in a checkpoint's `result_files` before relying on hosted copies. The terms allow deletion of tuned models at three days' notice.
- The spec requires `wandb.api_key` although the guide omits it, and it also accepts `mlflow` and `hf` integrations. Check the price in the console before starting a job.

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.