trl/pr_5617/en/_app/immutable/nodes/57.2be98aaa.js

import{s as xt,n as St,o as jt}from"../chunks/scheduler.7b731bd4.js";import{S as Ct,i as Ut,e as r,s as a,c as d,h as It,a as s,d as n,b as l,f as ee,g as p,j as m,k as te,l as i,m as o,n as c,t as _,o as u,p as g}from"../chunks/index.cc268345.js";import{C as $t,H as J,E as kt}from"../chunks/MermaidChart.svelte_svelte_type_style_lang.8cd01f78.js";import{D as ue}from"../chunks/Docstring.93078cb6.js";import{C as at}from"../chunks/CodeBlock.919b6e8d.js";function At(lt){let b,ge,ce,fe,N,ve,x,he,S,rt='Simple Self-Distillation (SSD) is described in <a href="https://huggingface.co/papers/2604.01193" rel="nofollow">Embarrassingly Simple Self-Distillation Improves Code Generation</a>.',Me,j,st="SSD samples completions from the model at a training-time temperature and truncation configuration, then fine-tunes on those raw, unverified samples with standard cross-entropy loss. It requires no reward model, verifier, teacher model, or reinforcement learning — only a set of problem prompts and the model itself.",be,C,it="In the current TRL implementation:",ye,U,mt='<li>the model generates completions at a specified training-time temperature (<code>temperature</code>) and truncation (<code>top_k</code>, <code>top_p</code>)</li> <li>the dataset only requires a <code>prompt</code> column</li> <li>training uses standard cross-entropy loss on the generated completions</li> <li>empty or single-line stub completions are filtered by default (<code>filter_empty=True</code>)</li> <li>the evaluation-time temperature and truncation are set independently at inference time</li> <li>vLLM can be used for faster generation via <code>use_vllm=True</code> (see <a href="vllm_integration">vLLM integration</a>)</li>',Te,I,we,$,Je,k,Ne,A,dt="Each example must provide:",xe,D,pt="<li><code>prompt</code>: the problem prompt (string or conversational format)</li>",Se,L,ct="No <code>privileged_context</code>, reward functions, or teacher model are needed.",je,E,Ce,B,_t="The paper identifies the following key hyperparameters:",Ue,F,ut="<li><strong><code>temperature</code></strong>: training-time sampling temperature (T_train). Higher values create more diverse samples but may include more noise. The paper uses T_train=0.6 with truncation.</li> <li><strong><code>top_k</code></strong> and <strong><code>top_p</code></strong>: training-time truncation parameters (rho_train). These suppress low-probability distractor tails during data synthesis.</li> <li><strong>T_eval</strong>: the evaluation-time decoding temperature is set independently at inference time. The paper shows that T_train and T_eval compose through an effective temperature T_eff = T_train * T_eval, with a broad optimal band.</li>",Ie,G,$e,X,gt='Use <a href="https://github.com/huggingface/trl/blob/main/trl/experimental/ssd/ssd.py" rel="nofollow"><code>trl/experimental/ssd/ssd.py</code></a> to launch SSD training from the command line. The script supports any causal LM from the Hub, custom local datasets via <code>--dataset_path</code>, and PEFT/LoRA via the standard <code>ModelConfig</code> flags.',ke,R,Ae,W,De,Z,ft='Use <a href="https://github.com/huggingface/trl/blob/main/trl/experimental/ssd/ssd_eval.py" rel="nofollow"><code>trl/experimental/ssd/ssd_eval.py</code></a> to evaluate a base model or an SSD-trained checkpoint on LiveCodeBench v6. The script uses vLLM for generation and LiveCodeBench’s official <code>codegen_metrics</code> for sandboxed <code>pass@k</code> scoring; default decoding parameters match Table 3 of the paper.',Le,V,Ee,z,Be,v,H,ze,ne,vt="Configuration class for <code>SSDTrainer</code>.",He,oe,ht=`Implements Simple Self-Distillation (SSD) from <a href="https://huggingface.co/papers/2604.01193" rel="nofollow"><em>Embarrassingly Simple Self-Distillation Improves Code
Generation</em></a>. SSD samples completions from the model at a training-time
temperature and truncation configuration, then fine-tunes on those raw, unverified samples with standard
cross-entropy loss.`,qe,ae,Mt=`The <code>temperature</code>, <code>top_k</code>, and <code>top_p</code> parameters control the training-time sampling configuration (T_train,
rho_train in the paper). The evaluation-time configuration (T_eval, rho_eval) is set independently at inference
time.`,Fe,q,Ge,f,Q,Qe,le,bt="Trainer for SSD-style on-policy self-distillation with cross-entropy loss.",Pe,re,yt=`SSD generates completions from the model at a specified training-time temperature and truncation configuration,
then fine-tunes on those raw, unverified samples using standard cross-entropy loss. The dataset only requires a
<code>prompt</code> column.`,Ye,y,P,Oe,se,Tt="Main training entry point.",Ke,M,Y,et,ie,wt="Will save the model, so you can reload it using <code>from_pretrained()</code>.",tt,me,Jt="Will only save from the main process.",nt,T,O,ot,de,Nt="Upload <code>self.model</code> and <code>self.processing_class</code> to the 🤗 model hub on the repo <code>self.args.hub_model_id</code>.",Xe,K,Re,_e,We;return N=new $t({props:{containerStyle:"float: right; margin-left: 10px; display: inline-flex; position: relative; z-index: 10;"}}),x=new J({props:{title:"SSD",local:"ssd",headingTag:"h1"}}),I=new J({props:{title:"Usage",local:"usage",headingTag:"h2"}}),$=new at({props:{code:"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",highlighted:`<span class="hljs-keyword">from</span> datasets <span class="hljs-keyword">import</span> Dataset

<span class="hljs-keyword">from</span> trl.experimental.ssd <span class="hljs-keyword">import</span> SSDConfig, SSDTrainer

dataset = Dataset.from_dict(
    {
        <span class="hljs-string">&quot;prompt&quot;</span>: [
            [{<span class="hljs-string">&quot;role&quot;</span>: <span class="hljs-string">&quot;user&quot;</span>, <span class="hljs-string">&quot;content&quot;</span>: <span class="hljs-string">&quot;Write a function to add two numbers.&quot;</span>}],
            [{<span class="hljs-string">&quot;role&quot;</span>: <span class="hljs-string">&quot;user&quot;</span>, <span class="hljs-string">&quot;content&quot;</span>: <span class="hljs-string">&quot;Write a function to check if a number is prime.&quot;</span>}],
        ],
    }
)

training_args = SSDConfig(
    output_dir=<span class="hljs-string">&quot;ssd-model&quot;</span>,
    temperature=<span class="hljs-number">0.6</span>,           <span class="hljs-comment"># T_train from the paper</span>
    top_k=<span class="hljs-number">20</span>,                  <span class="hljs-comment"># training-time top-k truncation</span>
    top_p=<span class="hljs-number">0.95</span>,                <span class="hljs-comment"># training-time top-p truncation</span>
    max_completion_length=<span class="hljs-number">65536</span>,
    learning_rate=<span class="hljs-number">5e-6</span>,
)

trainer = SSDTrainer(
    model=<span class="hljs-string">&quot;Qwen/Qwen3-4B-Instruct&quot;</span>,
    args=training_args,
    train_dataset=dataset,
)
trainer.train()`,wrap:!1}}),k=new J({props:{title:"Expected dataset columns",local:"expected-dataset-columns",headingTag:"h2"}}),E=new J({props:{title:"Key hyperparameters",local:"key-hyperparameters",headingTag:"h2"}}),G=new J({props:{title:"Example script",local:"example-script",headingTag:"h2"}}),R=new at({props:{code:"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",highlighted:`python trl/experimental/ssd/ssd.py \\
    --model_name_or_path Qwen/Qwen3-4B-Instruct-2507 \\
    --dataset_name microsoft/rStar-Coder \\
    --dataset_config seed_sft \\
    --prompt_column question \\
    --output_dir outputs/ssd-qwen3-4b \\
    --per_device_train_batch_size 1 \\
    --gradient_accumulation_steps 32 \\
    --learning_rate 5e-6 \\
    --lr_scheduler_type cosine \\
    --max_prompt_length 1024 \\
    --max_completion_length 65536 \\
    --temperature 1.6 \\
    --top_k 20 \\
    --top_p 0.8 \\
    --num_train_epochs 1 \\
    --bf16 \\
    --report_to trackio`,wrap:!1}}),W=new J({props:{title:"Evaluation on LiveCodeBench",local:"evaluation-on-livecodebench",headingTag:"h2"}}),V=new at({props:{code:"cHl0aG9uJTIwdHJsJTJGZXhwZXJpbWVudGFsJTJGc3NkJTJGc3NkX2V2YWwucHklMjAlNUMlMEElMjAlMjAlMjAlMjAtLW1vZGVsX25hbWVfb3JfcGF0aCUyMCUzQ3BhdGgtb3ItcmVwbyUzRSUyMCU1QyUwQSUyMCUyMCUyMCUyMC0tdGVtcGVyYXR1cmUlMjAxLjElMjAtLXRvcF9rJTIwMjAlMjAtLXRvcF9wJTIwMC44JTIwJTVDJTBBJTIwJTIwJTIwJTIwLS1uJTIwMSUyMCU1QyUwQSUyMCUyMCUyMCUyMC0tb3V0cHV0X2ZpbGUlMjBvdXRwdXRzJTJGbGNiX3Y2Lmpzb24=",highlighted:`python trl/experimental/ssd/ssd_eval.py \\
    --model_name_or_path &lt;path-or-repo&gt; \\
    --temperature 1.1 --top_k 20 --top_p 0.8 \\
    --n 1 \\
    --output_file outputs/lcb_v6.json`,wrap:!1}}),z=new J({props:{title:"SSDConfig",local:"trl.experimental.ssd.SSDConfig",headingTag:"h2"}}),H=new ue({props:{name:"class trl.experimental.ssd.SSDConfig",anchor:"trl.experimental.ssd.SSDConfig",parameters:[{name:"output_dir",val:": str | None = None"},{name:"per_device_train_batch_size",val:": int = 8"},{name:"num_train_epochs",val:": float = 3.0"},{name:"max_steps",val:": int = -1"},{name:"learning_rate",val:": float = 5e-05"},{name:"lr_scheduler_type",val:": transformers.trainer_utils.SchedulerType | str = 'linear'"},{name:"lr_scheduler_kwargs",val:": dict | str | None = None"},{name:"warmup_steps",val:": float = 0"},{name:"optim",val:": transformers.training_args.OptimizerNames | str = 'adamw_torch_fused'"},{name:"optim_args",val:": str | None = None"},{name:"weight_decay",val:": float = 0.0"},{name:"adam_beta1",val:": float = 0.9"},{name:"adam_beta2",val:": float = 0.999"},{name:"adam_epsilon",val:": float = 1e-08"},{name:"optim_target_modules",val:": None | str | list[str] = None"},{name:"gradient_accumulation_steps",val:": int = 1"},{name:"average_tokens_across_devices",val:": bool = True"},{name:"max_grad_norm",val:": float = 1.0"},{name:"label_smoothing_factor",val:": float = 0.0"},{name:"bf16",val:": bool | None = None"},{name:"fp16",val:": bool = False"},{name:"bf16_full_eval",val:": bool = False"},{name:"fp16_full_eval",val:": bool = False"},{name:"tf32",val:": bool | None = None"},{name:"gradient_checkpointing",val:": bool = True"},{name:"gradient_checkpointing_kwargs",val:": dict[str, typing.Any] | str | None = None"},{name:"torch_compile",val:": bool = False"},{name:"torch_compile_backend",val:": str | None = None"},{name:"torch_compile_mode",val:": str | None = None"},{name:"use_liger_kernel",val:": bool = False"},{name:"liger_kernel_config",val:": dict[str, bool] | None = None"},{name:"use_cache",val:": bool = False"},{name:"neftune_noise_alpha",val:": float | None = None"},{name:"torch_empty_cache_steps",val:": int | None = None"},{name:"auto_find_batch_size",val:": bool = False"},{name:"logging_strategy",val:": transformers.trainer_utils.IntervalStrategy | str = 'steps'"},{name:"logging_steps",val:": float = 10"},{name:"logging_first_step",val:": bool = False"},{name:"log_on_each_node",val:": bool = True"},{name:"logging_nan_inf_filter",val:": bool = True"},{name:"include_num_input_tokens_seen",val:": str | bool = 'no'"},{name:"log_level",val:": str = 'passive'"},{name:"log_level_replica",val:": str = 'warning'"},{name:"disable_tqdm",val:": bool | None = None"},{name:"report_to",val:": None | str | list[str] = 'none'"},{name:"run_name",val:": str | None = None"},{name:"project",val:": str = 'huggingface'"},{name:"trackio_space_id",val:": str | None = 'trackio'"},{name:"eval_strategy",val:": transformers.trainer_utils.IntervalStrategy | str = 'no'"},{name:"eval_steps",val:": float | None = None"},{name:"eval_delay",val:": float = 0"},{name:"per_device_eval_batch_size",val:": int = 8"},{name:"prediction_loss_only",val:": bool = False"},{name:"eval_on_start",val:": bool = False"},{name:"eval_do_concat_batches",val:": bool = True"},{name:"eval_use_gather_object",val:": bool = False"},{name:"eval_accumulation_steps",val:": int | None = None"},{name:"include_for_metrics",val:": list = <factory>"},{name:"batch_eval_metrics",val:": bool = False"},{name:"save_only_model",val:": bool = False"},{name:"save_strategy",val:": transformers.trainer_utils.SaveStrategy | str = 'steps'"},{name:"save_steps",val:": float = 500"},{name:"save_on_each_node",val:": bool = False"},{name:"save_total_limit",val:": int | None = None"},{name:"enable_jit_checkpoint",val:": bool = False"},{name:"push_to_hub",val:": bool = False"},{name:"hub_token",val:": str | None = None"},{name:"hub_private_repo",val:": bool | None = None"},{name:"hub_model_id",val:": str | None = None"},{name:"hub_strategy",val:": transformers.trainer_utils.HubStrategy | str = 'every_save'"},{name:"hub_always_push",val:": bool = False"},{name:"hub_revision",val:": str | None = None"},{name:"load_best_model_at_end",val:": bool = False"},{name:"metric_for_best_model",val:": str | None = None"},{name:"greater_is_better",val:": bool | None = None"},{name:"ignore_data_skip",val:": bool = False"},{name:"restore_callback_states_from_checkpoint",val:": bool = False"},{name:"full_determinism",val:": bool = False"},{name:"seed",val:": int = 42"},{name:"data_seed",val:": int | None = None"},{name:"use_cpu",val:": bool = False"},{name:"accelerator_config",val:": dict | str | None = None"},{name:"parallelism_config",val:": accelerate.parallelism_config.ParallelismConfig | None = None"},{name:"dataloader_drop_last",val:": bool = False"},{name:"dataloader_num_workers",val:": int = 0"},{name:"dataloader_pin_memory",val:": bool = True"},{name:"dataloader_persistent_workers",val:": bool = False"},{name:"dataloader_prefetch_factor",val:": int | None = None"},{name:"remove_unused_columns",val:": bool = True"},{name:"label_names",val:": list[str] | None = None"},{name:"train_sampling_strategy",val:": str = 'random'"},{name:"length_column_name",val:": str = 'length'"},{name:"ddp_find_unused_parameters",val:": bool | None = None"},{name:"ddp_bucket_cap_mb",val:": int | None = None"},{name:"ddp_broadcast_buffers",val:": bool | None = None"},{name:"ddp_backend",val:": str | None = None"},{name:"ddp_timeout",val:": int = 1800"},{name:"fsdp",val:": list[transformers.trainer_utils.FSDPOption] | str | None = None"},{name:"fsdp_config",val:": dict[str, typing.Any] | str | None = None"},{name:"deepspeed",val:": dict | str | None = None"},{name:"debug",val:": str | list[transformers.debug_utils.DebugOption] = ''"},{name:"skip_memory_metrics",val:": bool = True"},{name:"do_train",val:": bool = False"},{name:"do_eval",val:": bool = False"},{name:"do_predict",val:": bool = False"},{name:"resume_from_checkpoint",val:": str | None = None"},{name:"warmup_ratio",val:": float | None = None"},{name:"logging_dir",val:": str | None = None"},{name:"local_rank",val:": int = -1"},{name:"model_init_kwargs",val:": dict[str, typing.Any] | None = None"},{name:"max_prompt_length",val:": int | None = 512"},{name:"max_completion_length",val:": int | None = 256"},{name:"generation_batch_size",val:": int | None = None"},{name:"steps_per_generation",val:": int | None = None"},{name:"temperature",val:": float = 1.0"},{name:"top_k",val:": int = 0"},{name:"top_p",val:": float = 1.0"},{name:"min_p",val:": float | None = None"},{name:"repetition_penalty",val:": float = 1.0"},{name:"generation_kwargs",val:": dict[str, typing.Any] | None = None"},{name:"cache_implementation",val:": str | None = None"},{name:"chat_template_kwargs",val:": dict[str, typing.Any] | None = None"},{name:"use_vllm",val:": bool = False"},{name:"vllm_mode",val:": str = 'colocate'"},{name:"vllm_model_impl",val:": str = 'vllm'"},{name:"vllm_server_base_url",val:": str | None = None"},{name:"vllm_server_host",val:": str = '0.0.0.0'"},{name:"vllm_server_port",val:": int = 8000"},{name:"vllm_server_timeout",val:": float = 240.0"},{name:"vllm_group_port",val:": int = 51216"},{name:"vllm_tensor_parallel_size",val:": int = 1"},{name:"vllm_gpu_memory_utilization",val:": float = 0.3"},{name:"vllm_max_model_length",val:": int | None = None"},{name:"vllm_enable_sleep_mode",val:": bool = False"},{name:"disable_dropout",val:": bool = True"},{name:"filter_empty",val:": bool = True"},{name:"num_iterations",val:": int = 1"},{name:"shuffle_dataset",val:": bool = True"},{name:"ds3_gather_for_generation",val:": bool = True"}],source:"https://github.com/huggingface/trl/blob/vr_5617/trl/experimental/ssd/ssd_config.py#L24",parameterGroups:[{title:"Parameters that control generation and rollout reuse",parametersDescription:[{anchor:"trl.experimental.ssd.SSDConfig.model_init_kwargs",description:`<strong>model_init_kwargs</strong> (<code>dict[str, Any]</code>, <em>optional</em>) &#x2014;
Keyword arguments used when the <code>model</code> argument is passed as a string.`,name:"model_init_kwargs"},{anchor:"trl.experimental.ssd.SSDConfig.max_prompt_length",description:`<strong>max_prompt_length</strong> (<code>int</code> or <code>None</code>, <em>optional</em>, defaults to <code>512</code>) &#x2014;
Maximum prompt length. Longer prompts are truncated from the left.`,name:"max_prompt_length"},{anchor:"trl.experimental.ssd.SSDConfig.max_completion_length",description:`<strong>max_completion_length</strong> (<code>int</code> or <code>None</code>, <em>optional</em>, defaults to <code>256</code>) &#x2014;
Maximum generated completion length.`,name:"max_completion_length"},{anchor:"trl.experimental.ssd.SSDConfig.generation_batch_size",description:`<strong>generation_batch_size</strong> (<code>int</code> or <code>None</code>, <em>optional</em>) &#x2014;
Global batch size used for generation. Mutually exclusive with <code>steps_per_generation</code>.`,name:"generation_batch_size"},{anchor:"trl.experimental.ssd.SSDConfig.steps_per_generation",description:`<strong>steps_per_generation</strong> (<code>int</code> or <code>None</code>, <em>optional</em>) &#x2014;
Number of optimizer steps that reuse one generated batch. Mutually exclusive with <code>generation_batch_size</code>.`,name:"steps_per_generation"}]},{title:"Parameters that control sampling",parametersDescription:[{anchor:"trl.experimental.ssd.SSDConfig.temperature",description:`<strong>temperature</strong> (<code>float</code>, <em>optional</em>, defaults to <code>1.0</code>) &#x2014;
Sampling temperature (T_train in the paper).`,name:"temperature"},{anchor:"trl.experimental.ssd.SSDConfig.top_k",description:`<strong>top_k</strong> (<code>int</code>, <em>optional</em>, defaults to <code>0</code>) &#x2014;
Top-k sampling parameter. <code>0</code> disables top-k filtering.`,name:"top_k"},{anchor:"trl.experimental.ssd.SSDConfig.top_p",description:`<strong>top_p</strong> (<code>float</code>, <em>optional</em>, defaults to <code>1.0</code>) &#x2014;
Top-p (nucleus) sampling parameter.`,name:"top_p"},{anchor:"trl.experimental.ssd.SSDConfig.min_p",description:`<strong>min_p</strong> (<code>float</code> or <code>None</code>, <em>optional</em>) &#x2014;
Minimum token probability for sampling.`,name:"min_p"},{anchor:"trl.experimental.ssd.SSDConfig.repetition_penalty",description:`<strong>repetition_penalty</strong> (<code>float</code>, <em>optional</em>, defaults to <code>1.0</code>) &#x2014;
Repetition penalty used during generation.`,name:"repetition_penalty"},{anchor:"trl.experimental.ssd.SSDConfig.generation_kwargs",description:`<strong>generation_kwargs</strong> (<code>dict[str, Any]</code> or <code>None</code>, <em>optional</em>) &#x2014;
Extra generation kwargs passed to <code>GenerationConfig</code>.`,name:"generation_kwargs"}]},{title:"Parameters that control vLLM generation",parametersDescription:[{anchor:"trl.experimental.ssd.SSDConfig.use_vllm",description:`<strong>use_vllm</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) &#x2014;
Whether to use vLLM for generation instead of the training model.`,name:"use_vllm"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_mode",description:`<strong>vllm_mode</strong> (<code>str</code>, <em>optional</em>, defaults to <code>&quot;colocate&quot;</code>) &#x2014;
vLLM mode: <code>&quot;colocate&quot;</code> (shared GPU) or <code>&quot;server&quot;</code> (separate vLLM server).`,name:"vllm_mode"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_model_impl",description:`<strong>vllm_model_impl</strong> (<code>str</code>, <em>optional</em>, defaults to <code>&quot;vllm&quot;</code>) &#x2014;
Model implementation for vLLM: <code>&quot;vllm&quot;</code>, <code>&quot;transformers&quot;</code>, or <code>&quot;auto&quot;</code>.`,name:"vllm_model_impl"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_server_base_url",description:`<strong>vllm_server_base_url</strong> (<code>str</code> or <code>None</code>, <em>optional</em>) &#x2014;
Base URL for the vLLM server. If provided, <code>vllm_server_host</code> and <code>vllm_server_port</code> are ignored.`,name:"vllm_server_base_url"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_server_host",description:`<strong>vllm_server_host</strong> (<code>str</code>, <em>optional</em>, defaults to <code>&quot;0.0.0.0&quot;</code>) &#x2014;
Host of the vLLM server (server mode only).`,name:"vllm_server_host"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_server_port",description:`<strong>vllm_server_port</strong> (<code>int</code>, <em>optional</em>, defaults to <code>8000</code>) &#x2014;
Port of the vLLM server (server mode only).`,name:"vllm_server_port"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_server_timeout",description:`<strong>vllm_server_timeout</strong> (<code>float</code>, <em>optional</em>, defaults to <code>240.0</code>) &#x2014;
Timeout in seconds to wait for the vLLM server.`,name:"vllm_server_timeout"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_group_port",description:`<strong>vllm_group_port</strong> (<code>int</code>, <em>optional</em>, defaults to <code>51216</code>) &#x2014;
Port for the weight update group (server mode only).`,name:"vllm_group_port"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_tensor_parallel_size",description:`<strong>vllm_tensor_parallel_size</strong> (<code>int</code>, <em>optional</em>, defaults to <code>1</code>) &#x2014;
Tensor parallel size for colocated vLLM.`,name:"vllm_tensor_parallel_size"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_gpu_memory_utilization",description:`<strong>vllm_gpu_memory_utilization</strong> (<code>float</code>, <em>optional</em>, defaults to <code>0.3</code>) &#x2014;
GPU memory utilization ratio for colocated vLLM.`,name:"vllm_gpu_memory_utilization"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_max_model_length",description:`<strong>vllm_max_model_length</strong> (<code>int</code> or <code>None</code>, <em>optional</em>) &#x2014;
Model context length for vLLM. Inferred from model config if not set.`,name:"vllm_max_model_length"},{anchor:"trl.experimental.ssd.SSDConfig.vllm_enable_sleep_mode",description:`<strong>vllm_enable_sleep_mode</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>False</code>) &#x2014;
Whether to enable sleep mode for colocated vLLM engine.`,name:"vllm_enable_sleep_mode"}]},{title:"Parameters that control training behavior",parametersDescription:[{anchor:"trl.experimental.ssd.SSDConfig.disable_dropout",description:`<strong>disable_dropout</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) &#x2014;
Whether to disable dropout in the model during training.`,name:"disable_dropout"},{anchor:"trl.experimental.ssd.SSDConfig.filter_empty",description:`<strong>filter_empty</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) &#x2014;
Whether to filter out empty or single-line stub completions from the generated data.`,name:"filter_empty"},{anchor:"trl.experimental.ssd.SSDConfig.num_iterations",description:`<strong>num_iterations</strong> (<code>int</code>, <em>optional</em>, defaults to <code>1</code>) &#x2014;
Number of optimization iterations per generated batch.`,name:"num_iterations"},{anchor:"trl.experimental.ssd.SSDConfig.shuffle_dataset",description:`<strong>shuffle_dataset</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) &#x2014;
Whether to shuffle the training dataset.`,name:"shuffle_dataset"},{anchor:"trl.experimental.ssd.SSDConfig.ds3_gather_for_generation",description:`<strong>ds3_gather_for_generation</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) &#x2014;
Whether to gather ZeRO-3 weights for generation.`,name:"ds3_gather_for_generation"},{anchor:"trl.experimental.ssd.SSDConfig.cache_implementation",description:`<strong>cache_implementation</strong> (<code>str</code> or <code>None</code>, <em>optional</em>) &#x2014;
Cache implementation used by transformers generation.`,name:"cache_implementation"},{anchor:"trl.experimental.ssd.SSDConfig.chat_template_kwargs",description:`<strong>chat_template_kwargs</strong> (<code>dict[str, Any]</code> or <code>None</code>, <em>optional</em>) &#x2014;
Extra kwargs forwarded to chat template application.`,name:"chat_template_kwargs"}]}]}}),q=new J({props:{title:"SSDTrainer",local:"trl.experimental.ssd.SSDTrainer",headingTag:"h2"}}),Q=new ue({props:{name:"class trl.experimental.ssd.SSDTrainer",anchor:"trl.experimental.ssd.SSDTrainer",parameters:[{name:"model",val:": str | PreTrainedModel | nn.Module"},{name:"args",val:": SSDConfig | None = None"},{name:"train_dataset",val:": Dataset | IterableDataset | None = None"},{name:"eval_dataset",val:": Dataset | IterableDataset | dict[str, Dataset | IterableDataset] | None = None"},{name:"processing_class",val:": PreTrainedTokenizerBase | ProcessorMixin | None = None"},{name:"callbacks",val:": list[TrainerCallback] | None = None"},{name:"optimizers",val:": tuple[torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None] = (None, None)"},{name:"peft_config",val:": PeftConfig | None = None"}],source:"https://github.com/huggingface/trl/blob/vr_5617/trl/experimental/ssd/ssd_trainer.py#L74"}}),P=new ue({props:{name:"train",anchor:"trl.experimental.ssd.SSDTrainer.train",parameters:[{name:"resume_from_checkpoint",val:": str | bool | None = None"},{name:"trial",val:": optuna.Trial | dict[str, Any] | None = None"},{name:"ignore_keys_for_eval",val:": list[str] | None = None"}],parametersDescription:[{anchor:"trl.experimental.ssd.SSDTrainer.train.resume_from_checkpoint",description:`<strong>resume_from_checkpoint</strong> (<code>str</code> or <code>bool</code>, <em>optional</em>) &#x2014;
If a <code>str</code>, local path to a saved checkpoint as saved by a previous instance of <code>Trainer</code>. If a
<code>bool</code> and equals <code>True</code>, load the last checkpoint in <em>args.output_dir</em> as saved by a previous instance
of <code>Trainer</code>. If present, training will resume from the model/optimizer/scheduler states loaded here.`,name:"resume_from_checkpoint"},{anchor:"trl.experimental.ssd.SSDTrainer.train.trial",description:`<strong>trial</strong> (<code>optuna.Trial</code> or <code>dict[str, Any]</code>, <em>optional</em>) &#x2014;
The trial run or the hyperparameter dictionary for hyperparameter search.`,name:"trial"},{anchor:"trl.experimental.ssd.SSDTrainer.train.ignore_keys_for_eval",description:`<strong>ignore_keys_for_eval</strong> (<code>list[str]</code>, <em>optional</em>) &#x2014;
A list of keys in the output of your model (if it is a dictionary) that should be ignored when
gathering predictions for evaluation during the training.`,name:"ignore_keys_for_eval"}],source:"https://github.com/huggingface/trl/blob/vr_5617/transformers/trainer.py#L1323",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>Object containing the global step count, training loss, and metrics.</p>
`,returnType:`<script context="module">export const metadata = 'undefined';<\/script>


<p><code>~trainer_utils.TrainOutput</code></p>
`}}),Y=new ue({props:{name:"save_model",anchor:"trl.experimental.ssd.SSDTrainer.save_model",parameters:[{name:"output_dir",val:": str | None = None"},{name:"_internal_call",val:": bool = False"}],source:"https://github.com/huggingface/trl/blob/vr_5617/transformers/trainer.py#L3746"}}),O=new ue({props:{name:"push_to_hub",anchor:"trl.experimental.ssd.SSDTrainer.push_to_hub",parameters:[{name:"commit_message",val:": str | None = 'End of training'"},{name:"blocking",val:": bool = True"},{name:"token",val:": str | None = None"},{name:"revision",val:": str | None = None"},{name:"**kwargs",val:""}],parametersDescription:[{anchor:"trl.experimental.ssd.SSDTrainer.push_to_hub.commit_message",description:`<strong>commit_message</strong> (<code>str</code>, <em>optional</em>, defaults to <code>&quot;End of training&quot;</code>) &#x2014;
Message to commit while pushing.`,name:"commit_message"},{anchor:"trl.experimental.ssd.SSDTrainer.push_to_hub.blocking",description:`<strong>blocking</strong> (<code>bool</code>, <em>optional</em>, defaults to <code>True</code>) &#x2014;
Whether the function should return only when the <code>git push</code> has finished.`,name:"blocking"},{anchor:"trl.experimental.ssd.SSDTrainer.push_to_hub.token",description:`<strong>token</strong> (<code>str</code>, <em>optional</em>, defaults to <code>None</code>) &#x2014;
Token with write permission to overwrite Trainer&#x2019;s original args.`,name:"token"},{anchor:"trl.experimental.ssd.SSDTrainer.push_to_hub.revision",description:`<strong>revision</strong> (<code>str</code>, <em>optional</em>) &#x2014;
The git revision to commit from. Defaults to the head of the &#x201C;main&#x201D; branch.`,name:"revision"},{anchor:"trl.experimental.ssd.SSDTrainer.push_to_hub.kwargs",description:`<strong>kwargs</strong> (<code>dict[str, Any]</code>, <em>optional</em>) &#x2014;
Additional keyword arguments passed along to <code>~Trainer.create_model_card</code>.`,name:"kwargs"}],source:"https://github.com/huggingface/trl/blob/vr_5617/transformers/trainer.py#L3993",returnDescription:`<script context="module">export const metadata = 'undefined';<\/script>


<p>The URL of the repository where the model was pushed if <code>blocking=False</code>, or a <code>Future</code> object tracking the
progress of the commit if <code>blocking=True</code>.</p>
`}}),K=new kt({props:{source:"https://github.com/huggingface/trl/blob/main/docs/source/ssd_trainer.md"}}),{c(){b=r("meta"),ge=a(),ce=r("p"),fe=a(),d(N.$$.fragment),ve=a(),d(x.$$.fragment),he=a(),S=r("p"),S.innerHTML=rt,Me=a(),j=r("p"),j.textContent=st,be=a(),C=r("p"),C.textContent=it,ye=a(),U=r("ul"),U.innerHTML=mt,Te=a(),d(I.$$.fragment),we=a(),d($.$$.fragment),Je=a(),d(k.$$.fragment),Ne=a(),A=r("p"),A.textContent=dt,xe=a(),D=r("ul"),D.innerHTML=pt,Se=a(),L=r("p"),L.innerHTML=ct,je=a(),d(E.$$.fragment),Ce=a(),B=r("p"),B.textContent=_t,Ue=a(),F=r("ul"),F.innerHTML=ut,Ie=a(),d(G.$$.fragment),$e=a(),X=r("p"),X.innerHTML=gt,ke=a(),d(R.$$.fragment),Ae=a(),d(W.$$.fragment),De=a(),Z=r("p"),Z.innerHTML=ft,Le=a(),d(V.$$.fragment),Ee=a(),d(z.$$.fragment),Be=a(),v=r("div"),d(H.$$.fragment),ze=a(),ne=r("p"),ne.innerHTML=vt,He=a(),oe=r("p"),oe.innerHTML=ht,qe=a(),ae=r("p"),ae.innerHTML=Mt,Fe=a(),d(q.$$.fragment),Ge=a(),f=r("div"),d(Q.$$.fragment),Qe=a(),le=r("p"),le.textContent=bt,Pe=a(),re=r("p"),re.innerHTML=yt,Ye=a(),y=r("div"),d(P.$$.fragment),Oe=a(),se=r("p"),se.textContent=Tt,Ke=a(),M=r("div"),d(Y.$$.fragment),et=a(),ie=r("p"),ie.innerHTML=wt,tt=a(),me=r("p"),me.textContent=Jt,nt=a(),T=r("div"),d(O.$$.fragment),ot=a(),de=r("p"),de.innerHTML=Nt,Xe=a(),d(K.$$.fragment),Re=a(),_e=r("p"),this.h()},l(e){const t=It("svelte-u9bgzb",document.head);b=s(t,"META",{name:!0,content:!0}),t.forEach(n),ge=l(e),ce=s(e,"P",{}),ee(ce).forEach(n),fe=l(e),p(N.$$.fragment,e),ve=l(e),p(x.$$.fragment,e),he=l(e),S=s(e,"P",{"data-svelte-h":!0}),m(S)!=="svelte-13vb12j"&&(S.innerHTML=rt),Me=l(e),j=s(e,"P",{"data-svelte-h":!0}),m(j)!=="svelte-1h9gkbg"&&(j.textContent=st),be=l(e),C=s(e,"P",{"data-svelte-h":!0}),m(C)!=="svelte-1k81tbl"&&(C.textContent=it),ye=l(e),U=s(e,"UL",{"data-svelte-h":!0}),m(U)!=="svelte-c4byef"&&(U.innerHTML=mt),Te=l(e),p(I.$$.fragment,e),we=l(e),p($.$$.fragment,e),Je=l(e),p(k.$$.fragment,e),Ne=l(e),A=s(e,"P",{"data-svelte-h":!0}),m(A)!=="svelte-wjd8dj"&&(A.textContent=dt),xe=l(e),D=s(e,"UL",{"data-svelte-h":!0}),m(D)!=="svelte-181lg2c"&&(D.innerHTML=pt),Se=l(e),L=s(e,"P",{"data-svelte-h":!0}),m(L)!=="svelte-o4oon2"&&(L.innerHTML=ct),je=l(e),p(E.$$.fragment,e),Ce=l(e),B=s(e,"P",{"data-svelte-h":!0}),m(B)!=="svelte-1nk2czo"&&(B.textContent=_t),Ue=l(e),F=s(e,"UL",{"data-svelte-h":!0}),m(F)!=="svelte-1v4vr1w"&&(F.innerHTML=ut),Ie=l(e),p(G.$$.fragment,e),$e=l(e),X=s(e,"P",{"data-svelte-h":!0}),m(X)!=="svelte-m66uy7"&&(X.innerHTML=gt),ke=l(e),p(R.$$.fragment,e),Ae=l(e),p(W.$$.fragment,e),De=l(e),Z=s(e,"P",{"data-svelte-h":!0}),m(Z)!=="svelte-1hrcx8n"&&(Z.innerHTML=ft),Le=l(e),p(V.$$.fragment,e),Ee=l(e),p(z.$$.fragment,e),Be=l(e),v=s(e,"DIV",{class:!0});var w=ee(v);p(H.$$.fragment,w),ze=l(w),ne=s(w,"P",{"data-svelte-h":!0}),m(ne)!=="svelte-1nw37tj"&&(ne.innerHTML=vt),He=l(w),oe=s(w,"P",{"data-svelte-h":!0}),m(oe)!=="svelte-dlg1wh"&&(oe.innerHTML=ht),qe=l(w),ae=s(w,"P",{"data-svelte-h":!0}),m(ae)!=="svelte-f2mymo"&&(ae.innerHTML=Mt),w.forEach(n),Fe=l(e),p(q.$$.fragment,e),Ge=l(e),f=s(e,"DIV",{class:!0});var h=ee(f);p(Q.$$.fragment,h),Qe=l(h),le=s(h,"P",{"data-svelte-h":!0}),m(le)!=="svelte-657y8y"&&(le.textContent=bt),Pe=l(h),re=s(h,"P",{"data-svelte-h":!0}),m(re)!=="svelte-9i7sg5"&&(re.innerHTML=yt),Ye=l(h),y=s(h,"DIV",{class:!0});var Ze=ee(y);p(P.$$.fragment,Ze),Oe=l(Ze),se=s(Ze,"P",{"data-svelte-h":!0}),m(se)!=="svelte-1cilnet"&&(se.textContent=Tt),Ze.forEach(n),Ke=l(h),M=s(h,"DIV",{class:!0});var pe=ee(M);p(Y.$$.fragment,pe),et=l(pe),ie=s(pe,"P",{"data-svelte-h":!0}),m(ie)!=="svelte-r8h4ov"&&(ie.innerHTML=wt),tt=l(pe),me=s(pe,"P",{"data-svelte-h":!0}),m(me)!=="svelte-1e6bius"&&(me.textContent=Jt),pe.forEach(n),nt=l(h),T=s(h,"DIV",{class:!0});var Ve=ee(T);p(O.$$.fragment,Ve),ot=l(Ve),de=s(Ve,"P",{"data-svelte-h":!0}),m(de)!=="svelte-8tudwd"&&(de.innerHTML=Nt),Ve.forEach(n),h.forEach(n),Xe=l(e),p(K.$$.fragment,e),Re=l(e),_e=s(e,"P",{}),ee(_e).forEach(n),this.h()},h(){te(b,"name","hf:doc:metadata"),te(b,"content",Dt),te(v,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),te(y,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),te(M,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),te(T,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8"),te(f,"class","docstring border-l-2 border-t-2 pl-4 pt-3.5 border-gray-100 rounded-tl-xl mb-6 mt-8")},m(e,t){i(document.head,b),o(e,ge,t),o(e,ce,t),o(e,fe,t),c(N,e,t),o(e,ve,t),c(x,e,t),o(e,he,t),o(e,S,t),o(e,Me,t),o(e,j,t),o(e,be,t),o(e,C,t),o(e,ye,t),o(e,U,t),o(e,Te,t),c(I,e,t),o(e,we,t),c($,e,t),o(e,Je,t),c(k,e,t),o(e,Ne,t),o(e,A,t),o(e,xe,t),o(e,D,t),o(e,Se,t),o(e,L,t),o(e,je,t),c(E,e,t),o(e,Ce,t),o(e,B,t),o(e,Ue,t),o(e,F,t),o(e,Ie,t),c(G,e,t),o(e,$e,t),o(e,X,t),o(e,ke,t),c(R,e,t),o(e,Ae,t),c(W,e,t),o(e,De,t),o(e,Z,t),o(e,Le,t),c(V,e,t),o(e,Ee,t),c(z,e,t),o(e,Be,t),o(e,v,t),c(H,v,null),i(v,ze),i(v,ne),i(v,He),i(v,oe),i(v,qe),i(v,ae),o(e,Fe,t),c(q,e,t),o(e,Ge,t),o(e,f,t),c(Q,f,null),i(f,Qe),i(f,le),i(f,Pe),i(f,re),i(f,Ye),i(f,y),c(P,y,null),i(y,Oe),i(y,se),i(f,Ke),i(f,M),c(Y,M,null),i(M,et),i(M,ie),i(M,tt),i(M,me),i(f,nt),i(f,T),c(O,T,null),i(T,ot),i(T,de),o(e,Xe,t),c(K,e,t),o(e,Re,t),o(e,_e,t),We=!0},p:St,i(e){We||(_(N.$$.fragment,e),_(x.$$.fragment,e),_(I.$$.fragment,e),_($.$$.fragment,e),_(k.$$.fragment,e),_(E.$$.fragment,e),_(G.$$.fragment,e),_(R.$$.fragment,e),_(W.$$.fragment,e),_(V.$$.fragment,e),_(z.$$.fragment,e),_(H.$$.fragment,e),_(q.$$.fragment,e),_(Q.$$.fragment,e),_(P.$$.fragment,e),_(Y.$$.fragment,e),_(O.$$.fragment,e),_(K.$$.fragment,e),We=!0)},o(e){u(N.$$.fragment,e),u(x.$$.fragment,e),u(I.$$.fragment,e),u($.$$.fragment,e),u(k.$$.fragment,e),u(E.$$.fragment,e),u(G.$$.fragment,e),u(R.$$.fragment,e),u(W.$$.fragment,e),u(V.$$.fragment,e),u(z.$$.fragment,e),u(H.$$.fragment,e),u(q.$$.fragment,e),u(Q.$$.fragment,e),u(P.$$.fragment,e),u(Y.$$.fragment,e),u(O.$$.fragment,e),u(K.$$.fragment,e),We=!1},d(e){e&&(n(ge),n(ce),n(fe),n(ve),n(he),n(S),n(Me),n(j),n(be),n(C),n(ye),n(U),n(Te),n(we),n(Je),n(Ne),n(A),n(xe),n(D),n(Se),n(L),n(je),n(Ce),n(B),n(Ue),n(F),n(Ie),n($e),n(X),n(ke),n(Ae),n(De),n(Z),n(Le),n(Ee),n(Be),n(v),n(Fe),n(Ge),n(f),n(Xe),n(Re),n(_e)),n(b),g(N,e),g(x,e),g(I,e),g($,e),g(k,e),g(E,e),g(G,e),g(R,e),g(W,e),g(V,e),g(z,e),g(H),g(q,e),g(Q),g(P),g(Y),g(O),g(K,e)}}}const Dt='{"title":"SSD","local":"ssd","sections":[{"title":"Usage","local":"usage","sections":[],"depth":2},{"title":"Expected dataset columns","local":"expected-dataset-columns","sections":[],"depth":2},{"title":"Key hyperparameters","local":"key-hyperparameters","sections":[],"depth":2},{"title":"Example script","local":"example-script","sections":[],"depth":2},{"title":"Evaluation on LiveCodeBench","local":"evaluation-on-livecodebench","sections":[],"depth":2},{"title":"SSDConfig","local":"trl.experimental.ssd.SSDConfig","sections":[],"depth":2},{"title":"SSDTrainer","local":"trl.experimental.ssd.SSDTrainer","sections":[],"depth":2}],"depth":1}';function Lt(lt){return jt(()=>{new URLSearchParams(window.location.search).get("fw")}),[]}class Rt extends Ct{constructor(b){super(),Ut(this,b,Lt,At,xt,{})}}export{Rt as component};