Skip to content

Llama 3.1 Nemotron English-Hebrew tool-calling QLoRA adapter

Replace every REPLACE_FROM_VERIFIED_BUNDLE marker from the final local evidence before validation or publication. Do not hand-copy metric values from console output.

Built with Llama

This is an unmerged PEFT QLoRA adapter for nvidia/Llama-3.1-Nemotron-Nano-8B-v1 at base and tokenizer revision 54641c1611fcff44fa4865626462445e0a153fc7. It was trained on paired English and Hebrew single-tool-call examples under Sommelier's recorded v3 contract.

Use and redistribution are subject to the NVIDIA Open Model License and the Llama 3.1 Community License. Review THIRD_PARTY.md, LICENSE-NVIDIA-OPEN-MODEL.txt, LICENSE-LLAMA-3.1.txt, and NOTICE before use. The publisher requires exact copies of the reviewed project files and rejects abbreviated or edited terms. This repository distributes adapter weights, not merged base weights.

Immutable evidence identity

  • Adapter tree SHA-256: REPLACE_FROM_VERIFIED_BUNDLE
  • Final experiment_report.json SHA-256: REPLACE_FROM_VERIFIED_BUNDLE
  • Producer source revision: REPLACE_FROM_VERIFIED_BUNDLE
  • Hebrew paired dataset revision: REPLACE_FROM_VERIFIED_BUNDLE

The sommelier/ directory contains the resolved config, succeeded run/train manifests, passing release preflight, and final claim-gated sommelier.experiment_report.v2. The experiment report is authoritative for cohort sizes, marginal and exact matched-pair metrics, ordered pairing identity, paired-bootstrap intervals, claim decisions, tokenizer tax, training runtime, memory, storage, and inference telemetry.

The immutable Hebrew dataset revision named above is separately validated to contain the exact Phase-A config, its pre-provider translation run identity, and the finalized semantic review carrying the preregistered human reviewer's signed attestation. Signature verification establishes possession of that configured public key and integrity of the attested decisions; it does not establish native fluency or semantic correctness.

Its training-tax evidence separates the observed combined en+he formatted-row workload into an English-only arithmetic counterfactual and an additive Hebrew increment, with data/token ratios and combined-vs-English multipliers. The counterfactual is not a separately trained English-only runtime or accuracy arm.

REPLACE_FROM_VERIFIED_BUNDLE_WITH_RENDERED_CLAIM_SECTION

Limitations

The Hebrew data is machine-translated and selection-conditioned. Its bounded semantic review is non-native and does not establish native fluency or complete semantic correctness. Results apply only to the pinned dataset, prompts, parser, decoding, model, adapter tree, and hardware/runtime evidence. Sequential inference telemetry is not serving throughput, projected tokens are not billed tokens, the Hebrew increment is selection-conditioned on accepted translated rows, and no full-fine-tuning or currency saving is claimed without a matched measurement.