entrenar/finetune/
mod.rs

1//! Fine-tuning pipeline for code generation models
2//!
3//! This module implements SPEC-FT-001: Rust Test Generation Fine-Tuning Pipeline.
4//!
5//! # Features
6//!
7//! - CUDA auto-detection with fallback to CPU
8//! - QLoRA fine-tuning with 4-bit quantization
9//! - Popperian 100-point falsification QA
10//! - Scientific reproducibility protocol
11//!
12//! # References
13//!
14//! - Hu et al. (2021) "LoRA: Low-Rank Adaptation of Large Language Models"
15//! - Dettmers et al. (2023) "QLoRA: Efficient Finetuning of Quantized LLMs"
16//! - Popper (1959) "The Logic of Scientific Discovery"
17
18mod backward_graph;
19pub mod classification;
20pub mod classify_eval_report;
21pub mod classify_pipeline;
22pub mod classify_trainer;
23pub mod classify_tuner;
24mod corpus;
25pub mod data_parallel;
26mod device;
27pub mod distributed;
28mod eval;
29mod fused_lora_clip;
30pub mod gpu_backward_fallback;
31pub mod gradient_server;
32pub mod instruct_corpus;
33pub mod instruct_pipeline;
34pub mod instruct_trainer;
35pub mod linear_probe;
36pub mod multi_adapter_pipeline;
37mod popperian;
38mod reproducibility;
39pub mod ring_allreduce;
40pub mod training_plan;
41pub mod tune_searchers;
42pub mod wgpu_pipeline;
43pub mod worker_client;
44
45#[cfg(test)]
46mod tests;
47#[cfg(test)]
48mod tests_classification_contract_falsify;
49#[cfg(test)]
50mod tests_ssc_contract_falsify;
51
52pub use classification::{
53    bce_with_logits_loss, compute_class_weights, corpus_stats, cross_entropy_loss,
54    load_multi_label_corpus, load_safety_corpus, ClassWeightStrategy, ClassificationHead,
55    MultiLabelSafetySample, SafetyCorpusStats, SafetySample, TokenizedSample,
56};
57pub use classify_eval_report::{evaluate_checkpoint, ClassifyEvalReport, SSC_LABELS};
58pub use classify_pipeline::{
59    BatchResult, ClassifyConfig, ClassifyPipeline, DataStats, DiagSeverity, HyperparamDiagnostic,
60    HyperparamDiagnostics,
61};
62pub use classify_trainer::{ClassifyTrainer, EpochMetrics, TrainResult, TrainingConfig};
63pub use classify_tuner::{
64    default_classify_search_space, extract_trial_params, ClassifyTuner, SchedulerKind,
65    TrialSummary, TuneConfig, TuneResult, TuneScheduler, TuneSearcher, TuneStrategy,
66};
67pub use corpus::{CorpusStats, SampleMetadata, TestGenCorpus, TestGenSample};
68pub use data_parallel::{
69    average_gradients, has_non_finite, shard_samples, DataParallelCoordinator,
70};
71pub use device::{ComputeDevice, DeviceInfo};
72pub use distributed::{DistributedConfig, NodeRole, WireMessage};
73pub use eval::{
74    contains_tautology, count_test_functions, has_edge_case_tests, has_meaningful_assertions,
75    EvalMetrics, EvalResult, TestEvaluator,
76};
77pub use gradient_server::{
78    AllReduceResult, BlockAllReduceResult, GradientServer, NonBlockAllReduceResult,
79};
80pub use instruct_corpus::{
81    format_chat_prompt, instruct_corpus_stats, load_instruct_corpus, InstructCorpusStats,
82    InstructMetadata, InstructSample,
83};
84pub use instruct_pipeline::{
85    GenerateConfig, InstructBatchResult, InstructConfig, InstructPipeline, InstructStepResult,
86};
87pub use instruct_trainer::{
88    InstructEpochMetrics, InstructTrainResult, InstructTrainer, InstructTrainingConfig,
89};
90pub use linear_probe::{
91    binary_mcc, bootstrap_mcc_ci, check_ship_gate, compare_baselines, compute_confidence_scores,
92    evaluate as evaluate_classification, generalization_test, should_escalate, BaselineComparison,
93    BootstrapCI, ClassificationMetrics, ConfidenceScore, EscalationLevel, GeneralizationResult,
94    LinearProbe, MlpProbe, ShipGateResult,
95};
96pub use multi_adapter_pipeline::{
97    AdapterConfig, AdapterSchedule, AdapterSlot, MultiAdapterPipeline,
98};
99pub use popperian::{PopperianQA, QAGrade};
100pub use reproducibility::{ExperimentLock, ReproducibilityConfig};
101pub use ring_allreduce::{allreduce_pair, RingAllReduceWorker};
102pub use training_plan::{
103    execute_plan, plan as training_plan, ApplyConfig, CheckStatus, DataAudit, HyperparameterPlan,
104    ManualConfig, ModelInfo, PlanConfig, PlanIssue, PlanVerdict, PreFlightCheck, ResourceEstimate,
105    TrainingPlan, TrialPreview,
106};
107pub use worker_client::{
108    AveragedBlockResult, AveragedNonBlockResult, AveragedResult, ShardAssignment, WorkerClient,
109};
entrenar/finetune/mod.rs

entrenar/finetune/
mod.rs
