Skip to content

Provenir

The trust layer for model post-training — reproducible, evaluation-first orchestration for LLM fine-tuning and RL.

v0.3.0 · Apache-2.0 · Python ≥ 3.11 · 909 tests · zero breaking changes

Provenir coordinates every layer of the fine-tuning and reinforcement-learning lifecycle — raw data ingestion, quality filtering, prompt templating, SFT/DPO/GRPO training, verifiable-reward RL, eval-in-the-loop early stopping, LLM-as-judge preference generation, RLAIF iteration, adapter merging, benchmark evaluation, and production serving — in a single auditable, reproducible pipeline. On top of that it adds the trust primitives that raw-throughput engines leave out: RL observability, reward-hacking detection, contamination-safe evaluation, deterministic replay, and a signed Model Passport.

Why Provenir?

Every major fine-tuning framework optimises for raw throughput. Provenir optimises for trust:

  • Reproducibility — every run produces a content-addressed RunManifest that captures config hash, dataset hash, git SHA, and the full provenance chain. Any run can be reproduced exactly with provenir reproduce.
  • Evaluation-gated iteration — the EvalCallback and RL eval gate evaluate mid-training and halt or raise an alarm when quality regresses, before the model reaches production.
  • RL you can trust — an RL Flight Recorder detects training pathologies, and a reward-hacking detector catches the #1 RL failure mode before it wastes GPU budget.
  • No human labellers required — the RLAIF pipeline uses an LLM judge to generate preference pairs, run DPO, evaluate, and iterate autonomously.
  • Governance built-in — PII scanning, secret detection, append-only audit log, model card generation, and a signed Model Passport ship in the core package, not as plugins.

Feature Matrix

Axolotl TRL Unsloth verl Provenir
SFT / DPO / GRPO ✓ ✓ ✓ ✓ ✓
LoRA / QLoRA ✓ ✓ ✓ ✓ ✓
Reproducible manifests — — — — ✓
Eval-in-the-loop — partial — — ✓
LLM-as-Judge — — — — ✓
RLAIF (no human labels) — — — — ✓
Data flywheel — — — — ✓
RAG dataset generation — — — — ✓
Adapter merging (SLERP/TIES/DARE) — — — — ✓
Standard benchmark suite — — — — ✓
Semantic decontamination — — — — ✓
REST API — — — — ✓
Governance & audit log — — — — ✓
RL Flight Recorder — — — — ✓
Reward-hacking detection — — — — ✓
Verifiable-reward environments (RLVR) — — — — ✓
Contamination firewall — — — — ✓
Judge calibration — — — — ✓
Deterministic replay / lineage DAG — — — — ✓
Signed Model Passport — — — — ✓

New in v0.3 — The Trust Layer

v0.3.0 turns Provenir into the trust layer for model post-training. Four pillars:

  • A. RL Flight Recorder + reward-hacking detection — a black box for RL runs (provenir.observability) flags KL blowup/collapse, entropy collapse, length explosion, advantage collapse, reward-std collapse, reward spikes, and gradient explosion; the reward-hacking detector catches length inflation, format exploits, test tampering, verifier gaming, and proxy-reward divergence.
  • B. Contamination-safe trustworthy eval + judge calibration — a contamination firewall (provenir.eval.contamination, provenir.eval.canary) with 13-gram / embedding / exact overlap detection, MinHash at scale, and canary-tagged private eval vaults, plus DebiasedJudge and EnsembleJudge calibration.
  • C. Verifiable-reward environments + GRPO/DAPO/GSPO orchestration — hack-resistant RLVR verifiers (provenir.environments) and an RLOrchestrator (provenir.train.rl) that runs rollout → verify → reward → flight recorder → hacking detector → eval-gate over verl / TRL / Unsloth via backend-agnostic adapters.
  • D. Deterministic replay + lineage DAG + signed Model Passport — content-addressed environment fingerprint, lineage DAG, and ReplayEngine (provenir.provenance) plus a signed (HMAC-SHA256) Model Passport / BOM (provenir.governance.passport) with compliance risk flags. Maps to EU AI Act Article 12.

Drop the whole trust layer into any training loop with three lines:

import provenir

with provenir.track("my-run", dataset=train_ds) as run:
    for step, metrics in training_loop():
        run.log_step(metrics)
    run.record_eval("mmlu", score=0.71)

# run.manifest, run.flight_recorder, run.hacking_report, run.passport

See the Trust Layer guide for full examples.

Quick Install

pip install provenir                        # core
pip install "provenir[train]"               # + SFT/DPO/LoRA via TRL
pip install "provenir[all]"                 # everything

Architecture

┌──────────────────────────────────────────────────────────────────┐
│                      provenir CLI / REST API                      │
└──────────────┬───────────────────────────────────┬───────────────┘
               │                                   │
   ┌───────────▼──────────┐           ┌────────────▼──────────────┐
   │      Orchestration    │           │       Data Pipeline        │
   │  RunManifest          │           │  JsonlDataset              │
   │  RLAIFPipeline        │           │  QualityScorer             │
   │  DataFlywheel         │           │  CurriculumSampler         │
   │  HyperparamSweep      │           │  RAGDataGenerator          │
   └───────────┬──────────┘           │  PromptTemplates (6)       │
               │                      └───────────────────────────┘
   ┌───────────▼──────────────────────────────────────────────────┐
   │                       Training Backends                        │
   │         TRLBackend (SFT · DPO · LoRA · QLoRA)                │
   │         DistributedConfig (FSDP · DeepSpeed · DDP)           │
   └───────────┬──────────────────────────────────────────────────┘
               │
   ┌───────────▼──────────────────────────────────────────────────┐
   │                      Evaluation Layer                          │
   │  ExactMatch · TokenF1 · BLEU-4 · ROUGE-L · Wilson CI         │
   │  RAGEvaluator · BenchmarkEvaluator (MMLU/HellaSwag/ARC/…)    │
   │  LLM-as-Judge (Stub / Cached / Anthropic / OpenAI)           │
   │  EvalCallback (early stopping · regression gate)              │
   └───────────┬──────────────────────────────────────────────────┘
               │
   ┌───────────▼──────────────────────────────────────────────────┐
   │                      Governance Layer                          │
   │  AuditLogger · PIIScanner · SecretScanner · ModelCardGen      │
   │  SemanticDecontamination · AdapterRegistry · HubClient        │
   └──────────────────────────────────────────────────────────────┘