> ## Documentation Index > Fetch the complete documentation index at: https://docs.nugen.in/llms.txt > Use this file to discover all available pages before exploring further. # Start Auto-Align > Submit a target alignment score and let the system plan the run as a recursive self-improvement loop: each checkpoint is scored, and the score steers the next phase. The system analyses your data, base model and prior train-time alignment history to build a multi-phase alignment plan. Phases are sequenced and executed continuously. The system may apply gradient-free representation adjustments, supervised learning, reward-guided optimisation, or a combination, re-planning when the score trajectory demands it. Checkpoints are evaluated between phases, and the system revisits earlier phases if regressions surface. Every parameter defaults to `auto`. You can override any of them to constrain the search, and the system performs best when given room to explore. **Path Parameters:** - `alignment_id`: The alignment project to auto-align. Its base model, data and prior train-time alignment history warm-start the planning. **Request Body:** - `target_score` (required): Target quality on a 0-99 scale - `benchmark_id` (optional): Benchmark that measures the achieved score against the target. Defaults to the benchmark already on the project - `method`, `representations`, `sequencing` (optional): Constrain which optimisation families run and in what order - `max_compute_hours` (optional): Upper bound on GPU hours across all phases - `notify_on_phase_transition` (optional): Notify on each phase change - `config` (optional): Take control of any part of the run, for reinforcement learning or supervised (SFT) settings: loss, optimizer, LoRA adapter, reward, advantage, reference policy, environment, dataset, sampling, training, checkpointing, evaluation, rendering and logging. Pass only what you want to pin; anything omitted stays auto and redundant settings are discarded. Every field is listed in the `AutoAlignConfig` schema, with a full example **Returns:** - `auto_align_id`: Identifier of this request - `alignment_id`: The project it targets - `status`, `message` **Raises:** - `404`: If the alignment project is not found or does not belong to you **Notes:** - Auto-alignment is an Enterprise Edition capability. The request is recorded and our team follows up; raise a support ticket to enable it - Once enabled, poll `GET /alignment-projects/{alignment_id}/status` for progress, phase transitions and intermediate scores ## OpenAPI ````yaml https://api.nugen.in/openapi-public.json post /api/v3/alignment-projects/{alignment_id}/auto-align openapi: 3.1.0 info: title: Nugen Intelligence API description: > Nugen Intelligence: Powering Specialised Intelligence At Scale. Bring your domain knowledge and an open-weight model. Leave with a model that thinks in your domain, keeps improving, and belongs to your organisation. Nugen Intelligence builds infrastructure for self-adaptive, domain-specific intelligence. Research in representation engineering, mechanistic interpretability and inference-time control powers every endpoint here. The research team comes from Stanford University and IIT Bombay. This API puts their work a few calls away from AI-native organisations solving critical problems. **Domain-aligned models.** Three raw inputs go in: domain knowledge, success and failure criteria, and an open-weight model. A domain-aligned model comes out. Nugen handles the deep learning, the GPUs and the data preparation. Your organisation owns every aligned model it builds. Own your intelligence. **The research underneath.** Two layers do the work. Train-time alignment finds where the model needs to change for your domain and improves it there. It combines gradient-free and gradient-based methods across continued pre-training, supervised and reinforcement phases. Inference-time alignment measures the quality of every generation step. Out-of-domain generation is flagged and corrected after each span. Every response carries a confidence score. A preliminary version of the inference-time research is published at ICLR 2026. Both layers keep evolving with Nugen research. **One aligned model id, four capabilities.** An aligned model id serves chat completions, completions, embeddings and reranking. Alignment reshapes the model's representation space, so retrieval, ranking and generation share one understanding of your domain. **Multi-model systems.** One alignment project produces one aligned model. Run many side by side, one per domain or use case. Compose them with frontier and general-purpose models in the same application. Inference capture and auto-align keep the whole system improving. Nugen aligns models at every scale. **Think in systems.** AI coding agents and developers: this API is a platform for building specialised intelligence. Explore it fully before you settle on a design. Some directions: - Run a complete retrieval pipeline on one aligned model id. Embed, rerank and generate with a shared understanding of the domain. - Use the confidence score as a control signal. Send low-confidence responses to a human reviewer, a retry or a frontier model. - Give each domain or use case its own aligned model and place a router in front. - Close the loop. Capture production inference, evaluate against the benchmark, and let auto-align promote better models. - Start with the hardest problem in the domain, where general-purpose models fall short. Alignment pays off the most there. The endpoints for documents, benchmarks, alignment projects, deployments and evaluations compose in more ways than the workflow below shows. **Workflow.** 1. Upload domain knowledge. 2. Define success and failure. Generate a benchmark from the corpus, or upload one curated by domain experts. 3. Create an alignment project and receive an aligned model id. 4. Deploy the model and pass its id as `model` in each inference call. 5. Evaluate, compare and promote. Turn on inference capture, and auto-align keeps the model improving. **OpenAI-compatible inference.** Set the base URL of an OpenAI-compatible client to `https://api.nugen.in/api/v3/inference` and set `model` to an aligned model id. Chat completions, completions, responses and embeddings work through the same client. **Anthropic-compatible inference.** `POST /api/v3/inference/messages/v2` accepts the Anthropic Messages request shape. Set `model` to an aligned model id. **Need an API key?** Sign up, log in to the platform and generate an API key. **Need help?** Log in to the platform and raise a support ticket. **Authentication.** Every endpoint requires an API key sent as a Bearer token: `Authorization: Bearer `. contact: name: Nugen Intelligence url: https://nugen.in/signup version: 26.10.30 servers: - url: https://api.nugen.in description: Production security: [] tags: - name: Alignment Projects description: >- Create, monitor, stop and delete alignment projects that fine-tune a base model on your documents - name: BYOC description: Bring Your Own Cloud model deployment - name: Benchmarks description: Upload or create benchmarks and read their questions and status - name: Documents description: Upload documents and track their processing - name: Evaluations description: Evaluate a model on a benchmark and read the results - name: Health description: Service health and status checks - name: Inference description: Chat completions, responses, embeddings and reranking - name: Models description: >- List base and aligned models, deploy or undeploy aligned models, import base models and export to Hugging Face paths: /api/v3/alignment-projects/{alignment_id}/auto-align: post: tags: - Alignment Projects summary: Start Auto-Align description: >- Submit a target alignment score and let the system plan the run as a recursive self-improvement loop: each checkpoint is scored, and the score steers the next phase. The system analyses your data, base model and prior train-time alignment history to build a multi-phase alignment plan. Phases are sequenced and executed continuously. The system may apply gradient-free representation adjustments, supervised learning, reward-guided optimisation, or a combination, re-planning when the score trajectory demands it. Checkpoints are evaluated between phases, and the system revisits earlier phases if regressions surface. Every parameter defaults to `auto`. You can override any of them to constrain the search, and the system performs best when given room to explore. **Path Parameters:** - `alignment_id`: The alignment project to auto-align. Its base model, data and prior train-time alignment history warm-start the planning. **Request Body:** - `target_score` (required): Target quality on a 0-99 scale - `benchmark_id` (optional): Benchmark that measures the achieved score against the target. Defaults to the benchmark already on the project - `method`, `representations`, `sequencing` (optional): Constrain which optimisation families run and in what order - `max_compute_hours` (optional): Upper bound on GPU hours across all phases - `notify_on_phase_transition` (optional): Notify on each phase change - `config` (optional): Take control of any part of the run, for reinforcement learning or supervised (SFT) settings: loss, optimizer, LoRA adapter, reward, advantage, reference policy, environment, dataset, sampling, training, checkpointing, evaluation, rendering and logging. Pass only what you want to pin; anything omitted stays auto and redundant settings are discarded. Every field is listed in the `AutoAlignConfig` schema, with a full example **Returns:** - `auto_align_id`: Identifier of this request - `alignment_id`: The project it targets - `status`, `message` **Raises:** - `404`: If the alignment project is not found or does not belong to you **Notes:** - Auto-alignment is an Enterprise Edition capability. The request is recorded and our team follows up; raise a support ticket to enable it - Once enabled, poll `GET /alignment-projects/{alignment_id}/status` for progress, phase transitions and intermediate scores operationId: alignment_projects_auto_align parameters: - name: alignment_id in: path required: true schema: type: string title: Alignment Id description: ID of the alignment project, returned when it was created requestBody: required: true content: application/json: schema: $ref: '#/components/schemas/AlignmentProjectAutoAlignRequest' responses: '202': description: Auto-alignment request accepted and recorded. content: application/json: schema: $ref: '#/components/schemas/AlignmentProjectAutoAlignResponse' '401': description: Invalid API key content: application/json: schema: $ref: '#/components/schemas/ErrorResponse' '403': description: Authorization header missing or not a Bearer token content: application/json: schema: $ref: '#/components/schemas/ErrorResponse' '404': description: Resource not found content: application/json: schema: $ref: '#/components/schemas/ErrorResponse' '422': description: Validation Error content: application/json: schema: $ref: '#/components/schemas/HTTPValidationError' security: - ApiKeyAuth: [] components: schemas: AlignmentProjectAutoAlignRequest: properties: target_score: type: number maximum: 99 minimum: 0 title: Target Score description: >- Target alignment quality on a 0-99 scale. The system applies recursive self-improvement across phases until this score is reached or the compute budget is consumed. Targets above 85 typically cause the system to explore multiple optimisation strategies in sequence. examples: - 88 benchmark_id: anyOf: - type: string - type: 'null' title: Benchmark Id description: >- Benchmark used to measure the achieved score against the target. Defaults to the benchmark already on the alignment project. method: anyOf: - type: string - type: 'null' title: Method description: >- Which optimisation families the system may use. On auto it selects and sequences them from the score trajectory, and may begin with one approach, move to another, and revisit earlier ones. default: auto representations: anyOf: - type: string - type: 'null' title: Representations description: >- How internal model representations are handled before and during alignment. On auto the system profiles the base model and picks a strategy before any gradient-based phase begins. default: auto sequencing: anyOf: - items: type: string type: array - type: 'null' title: Sequencing description: >- Explicit phase ordering. When null the system orders phases itself and may interleave or revisit them. When given, it follows the order and still controls the hyperparameters within each phase. max_compute_hours: anyOf: - type: number - type: 'null' title: Max Compute Hours description: >- Upper bound on GPU hours across all phases, allocated by expected marginal score gain. When null the run continues until the target. notify_on_phase_transition: anyOf: - type: boolean - type: 'null' title: Notify On Phase Transition description: Notify when the system moves between alignment phases. default: true config: anyOf: - $ref: '#/components/schemas/AutoAlignConfig' - type: 'null' description: >- Take control of any part of the run, for reinforcement learning or supervised (SFT) settings: loss, optimizer, reward, advantage, reference policy, environment, dataset, sampling, training, adapter, checkpointing, evaluation, rendering and logging. Pass only what you want to pin; anything omitted stays auto, and redundant settings are discarded. examples: - adapter: alpha: 32 dropout: 0 init_seed: 7 mode: lora quantization_bits: 4 rank: 16 train_attn: true train_mlp: true train_unembed: true advantage: center: group_mean estimator: gae gae_lambda: 0.95 gamma: 0.99 normalize_std: true scope: per_token skip_zero_advantage_groups: true base_model: Qwen/Qwen2.5-0.5B-Instruct checkpointing: keep_last_n: 3 mode: training promote_final: true save_every_steps: 50 dataset: answer_field: responses file_type: jsonl format: chat prompt_field: instructions shuffle: true split: train streaming: true synthetic: enabled: false environment: fail_fast: false max_turns: 8 num_agents: 1 retry_on_failure: true sandbox: backend: local enabled: false tools: [] type: multi_turn evaluation: eval_batch_size: 64 eval_every_steps: 50 eval_group_size: 1 eval_temperature: 0 metrics: - reward - task_success logging: log_metrics: - reward - loss - kl - entropy - grad_norm wandb_project: alignment-id loss: cispo: eps_max: 6 custom_weights_field: token_weights dro: beta: 0.05 entropy_coef: 0.001 is_ratio_level: token kl_coef: 0.02 ppo: clip_high: 0.2 clip_low: 0.2 token_reduction: sum type: cispo z_loss_coef: 0.0001 optimizer: beta1: 0.9 beta2: 0.95 eps: 1.e-8 grad_clip_norm: 1 learning_rate: 0.000025 lr_schedule: cosine type: adamw warmup_steps: 20 weight_decay: 0 reference_policy: ema_decay: 0.999 enabled: true kl_target: 0.05 source: frozen_snapshot rendering: add_special_tokens: true batch_unit: tokens max_seq_len: 65536 packing: true truncation: right reward: clip: - -1 - 1 composite: task_success: 1 tool_format: 0.2 llm_judge: criteria: Were the right words used to reach the correct answer? judge_model: Qwen/Qwen2.5-0.5B-Instruct normalize: per_group_zscore type: llm_judge sampling: enable_thinking: true group_size: 8 max_prompt_tokens: 4096 max_seq_len: 131072 max_tokens: 1024 min_p: 0 prompt_caching: true prompt_groups_per_step: 64 seed: 12345 session_affinity_key: trajectory_id stop: - <|im_end|> temperature: 1 thinking_effort: high top_k: 40 top_p: 0.95 training: batch_size: 256 early_stopping: enabled: true metric: eval_loss patience: 3 epochs: 3 gradient_accumulation: 1 inner_optim_steps: 2 max_steps: 2000 mode: async rollout_training_overlap: true seed: 42 steps: 500 type: object required: - target_score title: AlignmentProjectAutoAlignRequest description: >- Target a score and let the system reach it through recursive self-improvement. Every field except target_score defaults to auto. The deep configuration blocks are accepted as given and stored verbatim, so a run can pin any detail without the public schema enumerating every knob. AlignmentProjectAutoAlignResponse: properties: auto_align_id: type: string title: Auto Align Id description: Identifier of the auto-alignment request. alignment_id: type: string title: Alignment Id description: Alignment project the request targets. status: type: string const: RECEIVED title: Status description: Request status. RECEIVED means the request was recorded. message: type: string title: Message description: Human-readable outcome of the request. type: object required: - auto_align_id - alignment_id - status - message title: AlignmentProjectAutoAlignResponse description: Acknowledgement of an auto-alignment request. ErrorResponse: description: Body of an error response. properties: detail: anyOf: - type: string - $ref: '#/components/schemas/ErrorDetail' description: >- Human readable explanation of the error, or an object with a message and the id of the resource concerned examples: - Model not found title: Detail required: - detail title: ErrorResponse type: object HTTPValidationError: properties: detail: items: $ref: '#/components/schemas/ValidationError' type: array title: Detail type: object title: HTTPValidationError description: Body of a request validation error. AutoAlignConfig: properties: base_model: anyOf: - type: string - type: 'null' title: Base Model description: Base model by Hugging Face id; defaults to the alignment project's loss: anyOf: - $ref: '#/components/schemas/AutoAlignLoss' - type: 'null' description: Loss function and its settings. optimizer: anyOf: - $ref: '#/components/schemas/AutoAlignOptimizer' - type: 'null' description: Optimiser and learning rate schedule. reward: anyOf: - $ref: '#/components/schemas/AutoAlignReward' - type: 'null' description: How each rollout is scored. advantage: anyOf: - $ref: '#/components/schemas/AutoAlignAdvantage' - type: 'null' description: How rewards become advantages. reference_policy: anyOf: - $ref: '#/components/schemas/AutoAlignReferencePolicy' - type: 'null' description: The reference policy the model is kept close to. environment: anyOf: - $ref: '#/components/schemas/AutoAlignEnvironment' - type: 'null' description: The environment rollouts run in. dataset: anyOf: - $ref: '#/components/schemas/AutoAlignDataset' - type: 'null' description: The data the run trains on. sampling: anyOf: - $ref: '#/components/schemas/AutoAlignSampling' - type: 'null' description: How rollouts are sampled. training: anyOf: - $ref: '#/components/schemas/AutoAlignTraining' - type: 'null' description: Training steps, batch size and stopping rules. adapter: anyOf: - $ref: '#/components/schemas/AutoAlignAdapter' - type: 'null' description: Adapter type and size. checkpointing: anyOf: - $ref: '#/components/schemas/AutoAlignCheckpointing' - type: 'null' description: When checkpoints are saved. evaluation: anyOf: - $ref: '#/components/schemas/AutoAlignEvaluation' - type: 'null' description: Evaluation during the run. rendering: anyOf: - $ref: '#/components/schemas/AutoAlignRendering' - type: 'null' description: How prompts and conversations are rendered. logging: anyOf: - $ref: '#/components/schemas/AutoAlignLogging' - type: 'null' description: What the run logs. additionalProperties: true type: object title: AutoAlignConfig description: >- Settings that pin any part of an auto-align run. Anything left out stays on auto. ErrorDetail: additionalProperties: true description: >- Structured error detail: a message, with the id of the resource it concerns as an extra key. properties: message: description: Human readable explanation of the error title: Message type: string required: - message title: ErrorDetail type: object ValidationError: properties: loc: items: anyOf: - type: string - type: integer type: array title: Location msg: type: string title: Message type: type: string title: Error Type type: object required: - loc - msg - type title: ValidationError description: One request validation problem. AutoAlignLoss: properties: type: anyOf: - type: string - type: 'null' title: Type description: >- Objective: cross_entropy for SFT, or ppo, cispo, dro for reinforcement learning ppo: anyOf: - additionalProperties: type: number type: object - type: 'null' title: Ppo description: PPO clip bounds, clip_low and clip_high cispo: anyOf: - additionalProperties: type: number type: object - type: 'null' title: Cispo description: CISPO settings, eps_max dro: anyOf: - additionalProperties: type: number type: object - type: 'null' title: Dro description: DRO settings, beta kl_coef: anyOf: - type: number - type: 'null' title: Kl Coef description: KL penalty against the reference policy entropy_coef: anyOf: - type: number - type: 'null' title: Entropy Coef description: Entropy bonus token_reduction: anyOf: - type: string - type: 'null' title: Token Reduction description: 'How token losses reduce: sum or mean' is_ratio_level: anyOf: - type: string - type: 'null' title: Is Ratio Level description: 'Importance ratio level: token or sequence' custom_loss_ref: anyOf: - type: string - type: 'null' title: Custom Loss Ref description: Reference to a custom loss custom_weights_field: anyOf: - type: string - type: 'null' title: Custom Weights Field description: Dataset field holding per-token weights z_loss_coef: anyOf: - type: number - type: 'null' title: Z Loss Coef description: Z-loss coefficient additionalProperties: true type: object title: AutoAlignLoss description: The training objective and its settings. AutoAlignOptimizer: properties: type: anyOf: - type: string - type: 'null' title: Type description: Optimizer, e.g. adamw learning_rate: anyOf: - type: number - type: 'null' title: Learning Rate description: Peak learning rate beta1: anyOf: - type: number - type: 'null' title: Beta1 description: Adam beta1 beta2: anyOf: - type: number - type: 'null' title: Beta2 description: Adam beta2 eps: anyOf: - type: number - type: 'null' title: Eps description: Adam epsilon weight_decay: anyOf: - type: number - type: 'null' title: Weight Decay description: Weight decay grad_clip_norm: anyOf: - type: number - type: 'null' title: Grad Clip Norm description: Global gradient norm clip lr_schedule: anyOf: - type: string - type: 'null' title: Lr Schedule description: Learning rate schedule, e.g. cosine or linear warmup_steps: anyOf: - type: integer - type: 'null' title: Warmup Steps description: Warmup steps warmup_ratio: anyOf: - type: number - type: 'null' title: Warmup Ratio description: Warmup as a fraction of steps, instead of warmup_steps lr_multiplier: anyOf: - type: number - type: 'null' title: Lr Multiplier description: Multiplier on the learning rate additionalProperties: true type: object title: AutoAlignOptimizer description: The optimiser and its learning rate schedule. AutoAlignReward: properties: type: anyOf: - type: string - type: 'null' title: Type description: >- Reward source: llm_judge, verifier, rubric, preference_model or composite verifier_ref: anyOf: - type: string - type: 'null' title: Verifier Ref description: Reference to a programmatic verifier llm_judge: anyOf: - additionalProperties: true type: object - type: 'null' title: Llm Judge description: 'Judge settings: judge_model and criteria' rubric_ref: anyOf: - type: string - type: 'null' title: Rubric Ref description: Reference to a scoring rubric preference_model_ref: anyOf: - type: string - type: 'null' title: Preference Model Ref description: Reference to a preference model composite: anyOf: - additionalProperties: type: number type: object - type: 'null' title: Composite description: Weights per reward component clip: anyOf: - items: type: number type: array - type: 'null' title: Clip description: Reward clip range, [low, high] normalize: anyOf: - type: string - type: 'null' title: Normalize description: Reward normalisation, e.g. per_group_zscore additionalProperties: true type: object title: AutoAlignReward description: How each rollout is scored. AutoAlignAdvantage: properties: estimator: anyOf: - type: string - type: 'null' title: Estimator description: Advantage estimator, e.g. gae or group_relative center: anyOf: - type: string - type: 'null' title: Center description: Baseline used to centre advantages, e.g. group_mean normalize_std: anyOf: - type: boolean - type: 'null' title: Normalize Std description: Divide advantages by their standard deviation skip_zero_advantage_groups: anyOf: - type: boolean - type: 'null' title: Skip Zero Advantage Groups description: Skip groups whose advantages are all zero gamma: anyOf: - type: number - type: 'null' title: Gamma description: Discount factor gae_lambda: anyOf: - type: number - type: 'null' title: Gae Lambda description: GAE lambda scope: anyOf: - type: string - type: 'null' title: Scope description: Per token or per sequence additionalProperties: true type: object title: AutoAlignAdvantage description: >- How rewards become advantages, including the estimator, baseline and discounting. AutoAlignReferencePolicy: properties: enabled: anyOf: - type: boolean - type: 'null' title: Enabled description: Keep a reference policy for the KL penalty source: anyOf: - type: string - type: 'null' title: Source description: frozen_snapshot or ema snapshot_path: anyOf: - type: string - type: 'null' title: Snapshot Path description: Snapshot to use as the reference ema_decay: anyOf: - type: number - type: 'null' title: Ema Decay description: EMA decay when source is ema kl_target: anyOf: - type: number - type: 'null' title: Kl Target description: Target KL for adaptive control additionalProperties: true type: object title: AutoAlignReferencePolicy description: The reference policy the KL penalty is measured against. AutoAlignEnvironment: properties: type: anyOf: - type: string - type: 'null' title: Type description: single_turn or multi_turn env_ref: anyOf: - type: string - type: 'null' title: Env Ref description: Reference to an environment max_turns: anyOf: - type: integer - type: 'null' title: Max Turns description: Maximum turns per episode tools: anyOf: - items: type: string type: array - type: 'null' title: Tools description: Tools available to the model num_agents: anyOf: - type: integer - type: 'null' title: Num Agents description: Agents per episode rollout_strategy: anyOf: - type: string - type: 'null' title: Rollout Strategy description: Rollout strategy sandbox: anyOf: - additionalProperties: true type: object - type: 'null' title: Sandbox description: 'Sandbox settings: enabled and backend' retry_on_failure: anyOf: - type: boolean - type: 'null' title: Retry On Failure description: Retry a failed rollout fail_fast: anyOf: - type: boolean - type: 'null' title: Fail Fast description: Stop the run on the first rollout failure additionalProperties: true type: object title: AutoAlignEnvironment description: >- The environment the rollouts run in, including turns, tools and sandboxing. AutoAlignDataset: properties: name: anyOf: - type: string - type: 'null' title: Name description: Dataset name; defaults to the alignment project's data split: anyOf: - type: string - type: 'null' title: Split description: Split to train on prompt_field: anyOf: - type: string - type: 'null' title: Prompt Field description: Field holding the prompt answer_field: anyOf: - type: string - type: 'null' title: Answer Field description: Field holding the reference answer completion_field: anyOf: - type: string - type: 'null' title: Completion Field description: Field holding a completion shuffle: anyOf: - type: boolean - type: 'null' title: Shuffle description: Shuffle examples streaming: anyOf: - type: boolean - type: 'null' title: Streaming description: Stream instead of loading in memory format: anyOf: - type: string - type: 'null' title: Format description: chat or completion file_type: anyOf: - type: string - type: 'null' title: File Type description: jsonl, json or parquet max_examples: anyOf: - type: integer - type: 'null' title: Max Examples description: Cap on examples used synthetic: anyOf: - additionalProperties: true type: object - type: 'null' title: Synthetic description: >- Synthetic training dataset: enabled, domain_description, label_list, num_examples additionalProperties: true type: object title: AutoAlignDataset description: The data the run trains on and how it is read. AutoAlignSampling: properties: group_size: anyOf: - type: integer - type: 'null' title: Group Size description: Completions sampled per prompt prompt_groups_per_step: anyOf: - type: integer - type: 'null' title: Prompt Groups Per Step description: Prompt groups per step max_tokens: anyOf: - type: integer - type: 'null' title: Max Tokens description: Maximum generated tokens max_prompt_tokens: anyOf: - type: integer - type: 'null' title: Max Prompt Tokens description: Maximum prompt tokens max_seq_len: anyOf: - type: integer - type: 'null' title: Max Seq Len description: Maximum sequence length temperature: anyOf: - type: number - type: 'null' title: Temperature description: Sampling temperature top_p: anyOf: - type: number - type: 'null' title: Top P description: Nucleus sampling top_k: anyOf: - type: integer - type: 'null' title: Top K description: Top-k sampling min_p: anyOf: - type: number - type: 'null' title: Min P description: Min-p sampling stop: anyOf: - items: type: string type: array - type: 'null' title: Stop description: Stop sequences seed: anyOf: - type: integer - type: 'null' title: Seed description: Sampling seed enable_thinking: anyOf: - type: boolean - type: 'null' title: Enable Thinking description: Allow a thinking phase thinking_effort: anyOf: - type: string - type: 'null' title: Thinking Effort description: low, medium or high prompt_caching: anyOf: - type: boolean - type: 'null' title: Prompt Caching description: Cache shared prompt prefixes session_affinity_key: anyOf: - type: string - type: 'null' title: Session Affinity Key description: Field that keeps a trajectory on one worker additionalProperties: true type: object title: AutoAlignSampling description: How completions are sampled for each prompt. AutoAlignTraining: properties: steps: anyOf: - type: integer - type: 'null' title: Steps description: Optimiser steps batch_size: anyOf: - type: integer - type: 'null' title: Batch Size description: Batch size gradient_accumulation: anyOf: - type: integer - type: 'null' title: Gradient Accumulation description: Gradient accumulation steps inner_optim_steps: anyOf: - type: integer - type: 'null' title: Inner Optim Steps description: Optimiser steps per batch mode: anyOf: - type: string - type: 'null' title: Mode description: sync or async rollout_training_overlap: anyOf: - type: boolean - type: 'null' title: Rollout Training Overlap description: Overlap rollouts with optimisation epochs: anyOf: - type: integer - type: 'null' title: Epochs description: Epochs max_steps: anyOf: - type: integer - type: 'null' title: Max Steps description: Hard cap on steps seed: anyOf: - type: integer - type: 'null' title: Seed description: Seed early_stopping: anyOf: - additionalProperties: true type: object - type: 'null' title: Early Stopping description: 'Early stopping: enabled, metric, patience' additionalProperties: true type: object title: AutoAlignTraining description: Training steps, batch size, mode and stopping rules. AutoAlignAdapter: properties: mode: anyOf: - type: string - type: 'null' title: Mode description: lora or full rank: anyOf: - type: integer - type: 'null' title: Rank description: LoRA rank alpha: anyOf: - type: integer - type: 'null' title: Alpha description: LoRA alpha dropout: anyOf: - type: number - type: 'null' title: Dropout description: LoRA dropout train_attn: anyOf: - type: boolean - type: 'null' title: Train Attn description: Adapt attention layers train_mlp: anyOf: - type: boolean - type: 'null' title: Train Mlp description: Adapt MLP layers train_unembed: anyOf: - type: boolean - type: 'null' title: Train Unembed description: Adapt the unembedding layer init_seed: anyOf: - type: integer - type: 'null' title: Init Seed description: Adapter initialisation seed quantization_bits: anyOf: - type: integer - type: 'null' title: Quantization Bits description: Base model quantisation, e.g. 4 additionalProperties: true type: object title: AutoAlignAdapter description: Adapter type and size for the run. AutoAlignCheckpointing: properties: save_every_steps: anyOf: - type: integer - type: 'null' title: Save Every Steps description: Save a checkpoint every N steps mode: anyOf: - type: string - type: 'null' title: Mode description: Checkpoint mode keep_last_n: anyOf: - type: integer - type: 'null' title: Keep Last N description: Checkpoints to keep resume_from: anyOf: - type: string - type: 'null' title: Resume From description: Checkpoint to resume from promote_final: anyOf: - type: boolean - type: 'null' title: Promote Final description: Promote the final checkpoint to the aligned model additionalProperties: true type: object title: AutoAlignCheckpointing description: When checkpoints are saved during the run. AutoAlignEvaluation: properties: eval_dataset: anyOf: - type: string - type: 'null' title: Eval Dataset description: Evaluation dataset eval_every_steps: anyOf: - type: integer - type: 'null' title: Eval Every Steps description: Evaluate every N steps eval_group_size: anyOf: - type: integer - type: 'null' title: Eval Group Size description: Samples per evaluation prompt eval_temperature: anyOf: - type: number - type: 'null' title: Eval Temperature description: Evaluation temperature metrics: anyOf: - items: type: string type: array - type: 'null' title: Metrics description: Metrics to track benchmark_ref: anyOf: - type: string - type: 'null' title: Benchmark Ref description: Benchmark to evaluate against eval_batch_size: anyOf: - type: integer - type: 'null' title: Eval Batch Size description: Evaluation batch size additionalProperties: true type: object title: AutoAlignEvaluation description: Evaluation during the run. AutoAlignRendering: properties: renderer_name: anyOf: - type: string - type: 'null' title: Renderer Name description: Renderer chat_template: anyOf: - type: string - type: 'null' title: Chat Template description: Chat template override add_special_tokens: anyOf: - type: boolean - type: 'null' title: Add Special Tokens description: Add special tokens max_seq_len: anyOf: - type: integer - type: 'null' title: Max Seq Len description: Maximum rendered length truncation: anyOf: - type: string - type: 'null' title: Truncation description: left or right packing: anyOf: - type: boolean - type: 'null' title: Packing description: Pack sequences batch_unit: anyOf: - type: string - type: 'null' title: Batch Unit description: tokens or examples additionalProperties: true type: object title: AutoAlignRendering description: How prompts and conversations are rendered for the model. AutoAlignLogging: properties: wandb_project: anyOf: - type: string - type: 'null' title: Wandb Project description: Weights and Biases project log_metrics: anyOf: - items: type: string type: array - type: 'null' title: Log Metrics description: Metrics to log wandb_key: anyOf: - type: string format: password writeOnly: true - type: 'null' title: Wandb Key description: 'Weights and Biases API key. Write only: never returned or logged' additionalProperties: true type: object title: AutoAlignLogging description: Where the run logs its progress. securitySchemes: ApiKeyAuth: type: http description: Your Nugen API key, sent as a Bearer token scheme: bearer bearerFormat: API key ```` This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.