VW(1)

📦

version 8.6.1.dfsg1 · 2 versions

8.6.1.dfsg1 · ubuntu jammy, ubuntu noble (latest)
9.11.2 · nixos 26.05

💅 VW(1) — Vowpal Wabbit 💅

okay bestie, this is the User Commands manual for vw, aka Vowpal Wabbit — the fast online learning tool. This is part 1 of 3, so don't panic, we're just getting started. I'm gonna walk you through every single flag like we're ranking them in a tier list. No skips, no cap. ✨

📛 NAME 📛

vw - Vowpal Wabbit -- fast online learning tool

✨ DESCRIPTION ✨

💖 VW options 💖

--ring_size arg — sets the size of the example ring 💍
--onethread — turns off the parse thread 🧵

🔄 Update options 🔄

-l, --learning_rate arg — sets the learning rate 📈
--power_t arg — the t power value ⚡
--decay_learning_rate arg — sets the decay factor for learning_rate between passes 📉
--initial_t arg — the initial t value 🌀
--feature_mask arg — uses an existing regressor to decide which parameters can be updated. If no initial_regressor is given, it's also used for initial weights. 🎭

⚖️ Weight options ⚖️

-i, --initial_regressor arg — initial regressor(s) 🧠
--initial_weight arg — sets every weight to an initial value of arg. ⚖️
--random_weights arg — makes initial weights random 🎲
--normal_weights arg — makes initial weights normal 🔔
--truncated_normal_weights arg — makes initial weights truncated normal ✂️
--sparse_weights — uses a sparse datastructure for weights 🕸️
--input_feature_regularizer arg — per feature regularization input file 📂

🧵 Parallelization options 🧵

--span_server arg — location of the server for setting up a spanning tree 🌳
--threads — enable multi-threading 🧵
--unique_id arg (=0) — unique id used for cluster parallel jobs 🆔
--total arg (=1) — total number of nodes used in cluster parallel job 🔢
--node arg (=0) — node number in cluster parallel job 📍

🩺 Diagnostic options 🩺

--version — version information 📦
-a, --audit — print weights of features 🔍
-P, --progress arg — progress update frequency. int: additive, float: multiplicative 📊
--quiet — don't output disgnostics and progress updates 🤫
-h, --help — look here: http://hunch.net/~vw/ and click on Tutorial. 🆘

🎲 Random Seed option 🎲

--random_seed arg — seed random number generator 🎲

🧬 Feature options 🧬

--hash arg — how to hash the features. Available options: strings, all #️⃣
--hash_seed arg (=0) — seed for hash function 🌱
--ignore arg — ignore namespaces beginning with character <arg> 🚫
--ignore_linear arg — ignore namespaces beginning with character <arg> for linear terms only 📏
--keep arg — keep namespaces beginning with character <arg> ✅
--redefine arg — redefine namespaces beginning with characters of string S as namespace N. <arg> shall be in form 'N:=S' where := is operator. Empty N or S are treated as default namespace. Use ':' as a wildcard in S. 🔄
-b, --bit_precision arg — number of bits in the feature table 🔢
--noconstant — don't add a constant feature 🚫
-C, --constant arg — set initial value of constant 🔢
--ngram arg — generate N grams. To generate N grams for a single namespace 'foo', arg should be fN. 🔤
--skips arg — generate skips in N grams. This in conjunction with the ngram tag can be used to generate generalized n-skip-k-gram. To generate n-skips for a single namespace 'foo', arg should be fN. ⏭️
--feature_limit arg — limit to N features. To apply to a single namespace 'foo', arg should be fN 🔢
--affix arg — generate prefixes/suffixes of features; argument '+2a,-3b,+1' means generate 2-char prefixes for namespace a, 3-char suffixes for b and 1 char prefixes for default namespace 🔗
--spelling arg — compute spelling features for a give namespace (use '_' for default namespace) 🔤
--dictionary arg — read a dictionary for additional features (arg either 'x:file' or just 'file') 📖
--dictionary_path arg — look in this directory for dictionaries; defaults to current directory or env{PATH} 📂
--interactions arg — create feature interactions of any level between namespaces. 🔀
--permutations — use permutations instead of combinations for feature interactions of same namespace. 🔄
--leave_duplicate_interactions — don't remove interactions with duplicate combinations of namespaces. For ex. this is a duplicate: '-q ab -q ba' and a lot more in '-q ::'. 🚫
-q, --quadratic arg — create and use quadratic features 🟦
--q: arg — : corresponds to a wildcard for all printable characters 🃏
--cubic arg — create and use cubic features 🟪

🧪 Example options 🧪

-t, --testonly — ignore label information and just test 🧪
--holdout_off — no holdout data in multiple passes 🚫
--holdout_period arg (=10) — holdout period for test only ⏳
--holdout_after arg — holdout after n training examples, default off (disables holdout_period) ⏸️
--early_terminate arg (=3) — specify the number of passes tolerated when holdout loss doesn't decrease before early termination 🛑
--passes arg — number of training passes 🔁
--initial_pass_length arg — initial number of examples per pass 📏
--examples arg — number of examples to parse 🔢
--min_prediction arg — smallest prediction to output 🔽
--max_prediction arg — largest prediction to output 🔼
--sort_features — turn this on to disregard order in which features have been defined. This will lead to smaller cache sizes 🗂️
--loss_function arg (=squared) — specify the loss function to be used, uses squared by default. Currently available ones are squared, classic, hinge, logistic, quantile and poisson. 📉
--quantile_tau arg (=0.5) — parameter \tau associated with Quantile loss. Defaults to 0.5 📊
--l1 arg — l_1 lambda 🧮
--l2 arg — l_2 lambda 🧮
--no_bias_regularization arg — no bias in regularization 🚫
--named_labels arg — use names for labels (multiclass, etc.) rather than integers, argument specified all possible labels, comma-sep, eg "--named_labels Noun,Verb,Adj,Punc" 🏷️

💾 Output model 💾

-f, --final_regressor arg — final regressor 🏁
--readable_model arg — output human-readable final regressor with numeric features 📄
--invert_hash arg — output human-readable final regressor with feature names. Computationally expensive. 💸
--save_resume — save extra state so learning can be resumed later with new data 💾
--preserve_performance_counters — reset performance counters when warmstarting 🔄
--save_per_pass — save the model after every pass over data 💾
--output_feature_regularizer_binary arg — per feature regularization output file 📦
--output_feature_regularizer_text arg — per feature regularization output file, in text 📄
--id arg — user supplied ID embedded into the final regressor 🆔

📤 Output options 📤

-p, --predictions arg — file to output predictions to 📤
-r, --raw_predictions arg — file to output unnormalized predictions to 📤

🔍 Audit Regressor 🔍

--audit_regressor arg — stores feature names and their regressor values. Same dataset must be used for both regressor training and this mode. 🔍

🎯 Search options 🎯

--search arg — use learning to search, argument=maximum action id or 0 for LDF 🔎
--search_task arg — the search task (use "--search_task list" to get a list of available tasks) 📋
--search_metatask arg — the search metatask (use "--search_metatask list" to get a list of available metatasks) 📋
--search_interpolation arg — at what level should interpolation happen? [*data|policy] 🎚️
--search_rollout arg — how should rollouts be executed? [policy|oracle|*mix_per_state|mix_per_roll|none] 🎬
--search_rollin arg — how should past trajectories be generated? [policy|oracle|*mix_per_state|mix_per_roll] 🎞️
--search_passes_per_policy arg (=1) — number of passes per policy (only valid for search_interpolation=policy) 🔁
--search_beta arg (=0.5) — interpolation rate for policies (only valid for search_interpolation=policy) 🎚️
--search_alpha arg (=1.00000001e-10) — annealed beta = 1-(1-alpha)^t (only valid for search_interpolation=data) 📉
--search_total_nb_policies arg — if we are going to train the policies through multiple separate calls to vw, we need to specify this parameter and tell vw how many policies are eventually going to be trained 🔢
--search_trained_nb_policies arg — the number of trained policies in a file 📁
--search_allowed_transitions arg — read file of allowed transitions [def: all transitions are allowed] 📜
--search_subsample_time arg — instead of training at all timesteps, use a subset. if value in (0,1), train on a random v%. if v>=1, train on precisely v steps per example, if v<=-1, use active learning ⏱️
--search_neighbor_features arg — copy features from neighboring lines. argument looks like: '-1:a,+2' meaning copy previous line namespace a and next next line from namespace _unnamed_, where ',' separates them 🔗
--search_rollout_num_steps arg — how many calls of "loss" before we stop really predicting on rollouts and switch to oracle (default means "infinite") 🔄
--search_history_length arg (=1) — some tasks allow you to specify how much history their depend on; specify that here 📜
--search_no_caching — turn off the built-in caching ability (makes things slower, but technically more safe) 🚫
--search_xv — train two separate policies, alternating prediction/learning 🔀
--search_perturb_oracle arg (=0) — perturb the oracle on rollin with this probability 🎲
--search_linear_ordering — insist on generating examples in linear order (def: hoopla permutation) 📏
--search_active_verify arg — verify that active learning is doing the right thing (arg = multiplier, should be = cost_range * range_c) ✅
--search_save_every_k_runs arg — save model every k runs 💾

🔁 Experience Replay 🔁

--replay_c arg — use experience replay at a specified level [b=classification/regression, m=multiclass, c=cost sensitive] with specified buffer size 🔁
--replay_c_count arg (=1) — how many times (in expectation) should each example be played (default: 1 = permuting) 🔢

🧪 Explore evaluation 🧪

--explore_eval — Evaluate explore_eval adf policies 🧪
--multiplier arg — Multiplier used to make all rejection sample probabilities <= 1. no cap, it just scales everything so those probs stay in check 💅

🎀 Make Multiclass into Contextual Bandit 🎀

--cbify arg — Convert multiclass on <k> classes into a contextual bandit problem. it's giving glow-up ✨
--cbify_cs — consume cost-sensitive classification examples instead of multiclass. lowkey a whole different vibe 🦋
--loss0 arg (=0) — loss for correct label. bestie, you want this low 💖
--loss1 arg (=1) — loss for incorrect label. not the vibe, but we move 😭

🦋 Contextual Bandit Exploration with Action Dependent Features 🦋

--cb_explore_adf — Online explore-exploit for a contextual bandit problem with multiline action dependent features. main character energy fr fr 👑
--first arg — tau-first exploration. firsties! 🙈
--epsilon arg — epsilon-greedy exploration. the classic, iconic 💅
--bag arg — bagging-based exploration. bag it up 🎀
--cover arg — Online cover based exploration. cover me in glory ✨
--psi arg (=1) — disagreement parameter for cover. it's giving disagreement, but make it fashion 🌸
--nounif — do not explore uniformly on zero-probability actions in cover. no uniform exploration for you 🚫
--softmax — softmax exploration. soft like a cloud ☁️
--regcb — RegCB-elim exploration. elimination style, periodt 💀
--regcbopt — RegCB optimistic exploration. optimistic queen 👑
--mellowness arg (=0.100000001) — RegCB mellowness parameter c_0. Default 0.1. stay mellow, bestie 🧘
--greedify — always update first policy once in bagging. greedy but make it cute 💖
--cb_min_cost arg (=0) — lower bound on cost. keep it low, low, low 🎶
--cb_max_cost arg (=1) — upper bound on cost. cap it at 1, no cap 🧢
--first_only — Only explore the first action in a tie-breaking event. first is the worst, second is the best? not here, first only 🥇
--lambda arg (=-1) — parameter for softmax. lambda, slay 💅

🌈 Contextual Bandit Exploration 🌈

--cb_explore arg — Online explore-exploit for a <k> action contextual bandit problem. explore and exploit, the ultimate duo 🦋
--first arg — tau-first exploration. tau-riffic! ✨
--epsilon arg (=0.0500000007) — epsilon-greedy exploration. a lil epsilon, a lot of drama 💅
--bag arg — bagging-based exploration. bag secured 👜
--cover arg — Online cover based exploration. cover girl, put the bass in your walk 🎶
--psi arg (=1) — disagreement parameter for cover. we disagree, respectfully 🙅

🌸 Multiworld Testing Options 🌸

--multiworld_test arg — Evaluate features as a policies. test it like it's hot 🔥
--learn arg — Do Contextual Bandit learning on <n> classes. learn, queen, learn 👑
--exclude_eval — Discard mwt policy features before learning. throw it away, it's not for you 🗑️

💅 Contextual Bandit with Action Dependent Features 💅

--cb_adf — Do Contextual Bandit learning with multiline action dependent features. adf stands for absolutely dazzling features ✨
--rank_all — Return actions sorted by score order. ranking, we love to see it 📊
--no_predict — Do not do a prediction when training. no predictions, just vibes 🚫
--cb_type arg (=ips) — contextual bandit method to use in {ips,dm,dr, mtr}. pick your fighter: ips, dm, dr, or mtr 🥊

👑 Contextual Bandit Options 👑

--cb arg — Use contextual bandit learning with <k> costs. contextual bandit, but make it costs 💸
--cb_type arg (=dr) — contextual bandit method to use in {ips,dm,dr}. the holy trinity: ips, dm, dr 🙏
--eval — Evaluate a policy rather than optimizing. just evaluate, don't optimize, periodt 💅

🎀 Cost Sensitive One Against All with Label Dependent Features 🎀

--csoaa_ldf arg — Use one-against-all multiclass learning with label dependent features. one against all, the ultimate showdown ⚔️
--ldf_override arg — Override singleline or multiline from csoaa_ldf or wap_ldf, eg if stored in file. override it like a boss 👑
--csoaa_rank — Return actions sorted by score order. rank it up 🎶
--probabilities — predict probabilites of all classes. probability? more like proba-BILL-ity 💅
--wap_ldf arg — Use weighted all-pairs multiclass learning with label dependent features. Specify singleline or multiline. weighted all-pairs, serving looks 💃

✨ Interact via elementwise multiplication ✨

--interact arg — Put weights on feature products from namespaces <n1> and <n2>. interact, don't just sit there 💬

💖 Cost Sensitive One Against All 💖

--csoaa arg — One-against-all multiclass with <k> costs. one against all, but make it cost-sensitive 💸

🦋 Cost-sensitive Active Learning 🦋

--cs_active arg — Cost-sensitive active learning with <k> costs. active learning, but cost-conscious 💅
--simulation — cost-sensitive active learning simulation mode. just a simulation, bestie 🎮
--baseline — cost-sensitive active learning baseline. baseline, the foundation of it all 🏗️
--domination — cost-sensitive active learning use domination. Default 1. dominate the game 👑
--mellowness arg (=0.100000001) — mellowness parameter c_0. Default 0.1. stay mellow, stay c_0 🌊
--range_c arg (=0.5) — parameter controlling the threshold for per-label cost uncertainty. Default 0.5. range_c, controlling the uncertainty 🎛️
--max_labels arg (=18446744073709551615) — maximum number of label queries. max labels, the limit does not exist 🚀
--min_labels arg (=18446744073709551615) — minimum number of label queries. min labels, also huge, iykyk 🤯
--cost_max arg (=1) — cost upper bound. Default 1. cost max, cap it 🧢
--cost_min arg (=0) — cost lower bound. Default 0. cost min, floor it 🚗
--csa_debug — print debug stuff for cs_active. debug, debug, debug 🐛

🌈 Multilabel One Against All 🌈

--multilabel_oaa arg — One-against-all multilabel with <k> labels. multilabel, multi-slay 💅

🌸 importance weight classes 🌸

--classweight arg — importance weight multiplier for class. weight it up, make it important 💪

💅 Recall Tree 💅

--recall_tree arg — Use online tree for multiclass. tree time, bestie 🌳
--max_candidates arg — maximum number of labels per leaf in the tree. candidates, but make it max 🍃
--bern_hyper arg (=1) — recall tree depth penalty. bern_hyper, the depth penalty we deserve 💅
--max_depth arg — maximum depth of the tree, default log_2 (#classes). depth so deep, log_2 of classes 🤿
--node_only arg (=0) — only use node features, not full path features. node only, keep it simple 🎯
--randomized_routing arg (=0) — randomized routing. randomize the route, live a little 🗺️

👑 Logarithmic Time Multiclass Tree 👑

--log_multi arg — Use online tree for multiclass. log time, but make it tree 🌲
--no_progress — disable progressive validation. no progress? relatable 😭
--swap_resistance arg (=4) — higher = more resistance to swap, default=4. resist the swap, stay strong 💪

🎀 Error Correcting Tournament Options 🎀

--ect arg — Error correcting tournament with <k> labels. tournament arc, but for errors 🏆
--error arg (=0) — errors allowed by ECT. errors allowed, we love a forgiving queen 👑

✨ Boosting ✨

--boosting arg — Online boosting with <N> weak learners. boost it, bestie 🚀
--gamma arg (=0.100000001) — weak learner's edge (=0.1), used only by online BBM. gamma, the edge you need 🔪
--alg arg (=BBM) — specify the boosting algorithm: BBM (default), logistic (AdaBoost.OL.W), adaptive (AdaBoost.OL). pick your algorithm: BBM, logistic, or adaptive 🎲

💖 One Against All Options 💖

--oaa arg — One-against-all multiclass with <k> labels. oaa, the original one-against-all 💅
--oaa_subsample arg — subsample this number of negative examples when learning. subsample the negatives, we love a sample 🎲
--probabilities — predict probabilites of all classes. probabilities, but spelled with an extra i for flavor 💖
--scores — output raw scores per class. raw scores, no filter needed 🔥

🦋 Top K 🦋

--top arg — top k recommendation. top k, the only k that matters 👑

🌈 Experience Replay 🌈

--replay_m arg — use experience replay at a specified level [b=classification/regression, m=multiclass, c=cost sensitive] with specified buffer size. replay it like you mean it 🔁
--replay_m_count arg (=1) — how many times (in expectation) should each example be played (default: 1 = permuting). play it again, sam 🎵

🌸 Binary loss 🌸

--binary — report loss as binary classification on -1,1. binary loss, it's giving -1 or 1 💅

💅 Bootstrap 💅

--bootstrap arg — k-way bootstrap by online importance resampling. bootstrap, but make it online 👢
--bs_type arg — prediction type {mean,vote}. mean or vote? you decide 🗳️

👑 scorer options 👑

--link arg (=identity) — Specify the link function: identity, logistic, glf1 or poisson. link up with identity, logistic, glf1, or poisson 🔗

🎀 Stagewise polynomial options 🎀

--stage_poly — use stagewise polynomial feature learning. stagewise polynomial, serving features 💃
--sched_exponent arg (=1) — exponent controlling quantity of included features. exponent, but make it scheduling 📈
--batch_sz arg (=1000) — multiplier on batch size before including more features. batch size, multiplied by 1000, no cap 🧢
--batch_sz_no_doubling — batch_sz does not double. no doubling, we stay steady 🚫

✨ Low Rank Quadratics FA ✨

--lrqfa arg — use low rank quadratic features with field aware weights. low rank, high impact 💅

💖 Low Rank Quadratics 💖

--lrq arg — use low rank quadratic features. lrq, the low rank quadratic queen 👑
--lrqdropout — use dropout training for low rank quadratic features. dropout, but make it low rank 🎓

🦋 Autolink 🦋

--autolink arg — create link function with polynomial d. autolink, link it up automatically 🔗

🌈 Marginal 🌈

--marginal arg — substitute marginal label estimates for ids. marginal, but make it estimates 💅
--initial_denominator arg (=1) — initial denominator. denominator starts at 1, slay ➗
--initial_numerator arg (=0.5) — initial numerator. numerator starts at 0.5, half the drama 🎭
--compete — enable competition with marginal features. compete, bestie, compete 🏁
--update_before_learn arg (=0) — update marginal values before learning. update before learn, always ahead 🔄
--unweighted_marginals arg (=0) — ignore importance weights when computing marginals. ignore the weights, just the marginals 💅
--decay arg (=0) — decay multiplier per event (1e-3 for example). decay, decay, decay (like my will to live) 💀

🌸 Matrix Factorization Reduction 🌸

--new_mf arg — rank for reduction-based matrix factorization. matrix factorization, but make it reduced 🧮

💅 Neural Network 💅

--nn arg — Sigmoidal feedforward network with <k> hidden units. neural network, the main character 🧠
--inpass — Train or test sigmoidal feedforward network with input passthrough. input passthrough, pass it through 🚪
--multitask — Share hidden layer across all reduced tasks. multitask, share the layer, share the love 💕
--dropout — Train or test sigmoidal feedforward network using dropout. dropout, but for neural nets 🎓
--meanfield — Train or test sigmoidal feedforward network using mean field. meanfield, the meanest field 🌾

👑 Confidence 👑

--confidence — Get confidence for binary predictions. confidence, get you some 💅
--confidence_after_training — Confidence after training. confidence after training, the afterglow ✨

🎀 Active Learning with Cover 🎀

--active_cover — enable active learning with cover. active cover, cover me in learning 📚
--mellowness arg (=8) — active learning mellowness parameter c_0. Default 8. mellow to the max 🌊
--alpha arg (=1) — active learning variance upper bound parameter alpha. Default 1. alpha, the upper bound queen 👑
--beta_scale arg (=3.1622777) — active learning variance upper bound parameter beta_scale. Default sqrt(10). beta_scale, sqrt(10) vibes 🧮
--cover arg (=12) — cover size. Default 12. cover size 12, a dozen of slay 🥚
--oracular — Use Oracular-CAL style query or not. Default false. oracular, but only if you want 👁️

✨ Active Learning ✨

--active — enable active learning. active learning, engage 🚀
--simulation — active learning simulation mode. simulation mode, pretend it's real 🎮
--mellowness arg (=8) — active learning mellowness parameter c_0. Default 8. mellow again, because why not 🧘

💖 Experience Replay 💖

--replay_b arg — use experience replay at a specified level [b=classification/regression, m=multiclass, c=cost sensitive] with specified buffer size. replay_b, the b stands for bestie 💖
--replay_b_count arg (=1) — how many times (in expectation) should each example be played (default: 1 = permuting). play it again, and again, and again 🔁

🦋 Baseline options 🦋

--baseline — Learn an additive baseline (from constant features) and a residual separately in regression. baseline, the foundation of your regression 🏗️
--lr_multiplier arg — learning rate multiplier for baseline model. learning rate, but multiplied 💅
--global_only — use separate example with only global constant for baseline predictions. global only, keep it universal 🌍
--check_enabled — only use baseline when the example contains enabled flag. check enabled, only if you're enabled ✅

🌈 OjaNewton options 🌈

--OjaNewton — Online Newton with Oja's Sketch. OjaNewton, the sketchy Newton 🎨
--sketch_size arg (=10) — size of sketch. sketch size 10, a perfect ten 💅
--epoch_size arg (=1) — size of epoch. epoch size 1, one epoch at a time 🕐
--alpha arg (=1) — mutiplicative constant for indentiy. alpha, mutiplicative and indentiy (typos included) 💖
--alpha_inverse arg — one over alpha, similar to learning rate. one over alpha, like a learning rate but cuter 🥰
--learning_rate_cnt arg (=2) — constant for the learning rate 1/t. learning rate constant, 1/t, iconic 💅
--normalize arg (=1) — normalize the features or not. normalize it, or don't, you do you 💁
--random_init arg (=1) — do you want Oja's initialization to be randomized or not? defaults to 1 (so yes, randomizing 💫).

🌈 LBFGS and Conjugate Gradient options:

--conjugate_gradient — roll with conjugate-gradient-based optimization.
--bfgs — roll with bfgs optimization instead.
--hessian_on — actually use the second derivative in the line search, bestie.
--mem arg (=15) — memory in bfgs, default 15.
--termination arg (=0.00100000005) — the termination threshold. that's the exact default, no rounding, we don't do that here.

🧚 Latent Dirichlet Allocation:

--lda arg — run lda with <int> topics.
--lda_alpha arg (=0.100000001) — prior on the sparsity of per-document topic weights, default 0.100000001.
--lda_rho arg (=0.100000001) — prior on the sparsity of topic distributions, default 0.100000001.
--lda_D arg (=10000) — number of documents, default 10000.
--lda_epsilon arg (=0.00100000005) — loop convergence threshold, default 0.00100000005.
--minibatch arg (=1) — minibatch size, for LDA. default 1.
--math-mode arg (=0) — math mode: simd, accuracy, fast-approx. default 0.
--metrics arg (=0) — compute metrics. default 0.

🎀 Noop Learner:

--noop — do no learning. that's it. that's the whole vibe. do nothing, slay ✨

💅 Print psuedolearner:

okay the original really does spell it psuedolearner and i am not about to fix the source's spelling, i'm just the messenger 💀

--print — print examples.

👑 Gradient Descent Matrix Factorization:

--rank arg — rank for matrix factorization.

🌸 Network sending:

--sendto arg — send examples to <host>.

🦋 Stochastic Variance Reduced Gradient:

--svrg — Streaming Stochastic Variance Reduced Gradient.
--stage_size arg (=1) — number of passes per SVRG stage, default 1.

🔥 Follow the Regularized Leader:

--ftrl — FTRL: Follow the Proximal Regularized Leader.
--ftrl_alpha arg (=0.00499999989) — learning rate for FTRL optimization, default 0.00499999989.
--ftrl_beta arg (=0.100000001) — FTRL beta parameter, default 0.100000001.
--pistol — FTRL: Parameter-free Stochastic Learning.
--ftrl_alpha arg (=1) — learning rate for FTRL optimization, default 1.
--ftrl_beta arg (=0.5) — FTRL beta parameter, default 0.5.

not me noticing --ftrl_alpha and --ftrl_beta show up twice in this list with completely different defaults 🙈 i'm reporting exactly what the page says, both entries, both defaults, no merging, no editing, periodt.

⭐ Kernel SVM:

--ksvm — kernel svm.
--reprocess arg (=1) — number of reprocess steps for LASVM, default 1.
--pool_greedy — use greedy selection on mini pools.
--para_active — do parallel active learning.
--pool_size arg (=1) — size of pools for active learning, default 1.
--subsample arg (=1) — number of items to subsample from the pool, default 1.
--kernel arg (=linear) — type of kernel (rbf or linear (default)).
--bandwidth arg (=1) — bandwidth of rbf kernel, default 1.
--degree arg (=2) — degree of poly kernel, default 2.
--lambda arg — saving regularization for test time.

💖 Gradient Descent options:

--sgd — use regular stochastic gradient descent update.
--adaptive — use adaptive, individual learning rates.
--adax — use adaptive learning rates with x^2 instead of g^2x^2.
--invariant — use safe/importance aware updates.
--normalized — use per feature normalized updates.
--sparse_l2 arg (=0) — use per feature normalized updates. default 0.
--l1_state arg (=0) — use per feature normalized updates. default 0.
--l2_state arg (=1) — use per feature normalized updates. default 1.

yep, the last four all say "use per feature normalized updates" in the original. i'm not the one who wrote that, i'm just the one who's telling you 😭 verbatim, bestie.

🌈 Input options:

-d [ --data ] arg — Example Set.
--daemon — persistent daemon mode on port 26542.
--foreground — in persistent daemon mode, do not run in the background.
--port arg — port to listen on; use 0 to pick an unused port.
--num_children arg — number of children for persistent daemon mode.
--pid_file arg — write pid file in persistent daemon mode.
--port_file arg — write port used in persistent daemon mode.
-c [ --cache ] — use a cache. the default is <data>.cache.
--cache_file arg — the location(s) of cache_file.
--json — enable JSON parsing.
--dsjson — enable Decision Service JSON parsing.
-k [ --kill_cache ] — do not reuse an existing cache: create a new one always.
--compressed — use gzip format whenever possible. if a cache file is being created, this option creates a compressed cache file. a mixture of raw-text & compressed inputs are supported with autodetection.
--no_stdin — do not default to reading from stdin.

and that's the whole page, bestie, you made it to the end 🫶 this is from vw 8.6.1, dated December 2020, and it's the VW(1) manual. go forth and train something iconic 💅✨

📦 where this page comes from

ubuntu jammy: vowpal-wabbit 8.6.1.dfsg1-1build3 · BSD-3
ubuntu noble: vowpal-wabbit 8.6.1.dfsg1-1build3 · BSD-3
💖 girlified by deepseek-ai/DeepSeek-V4.1-Flash on 2026-10-18: a language model's rewrite of the original, under its terms