jeongseokoh/Llama-SPEED-archive-deltagate-20260908

Llama SPEED research archive: deltagate

Built with Llama. This is a collection of historical SPEED experiments, including pilot/smoke checkpoints, for reproducibility of phase-asymmetric inference research. These are experimental artifacts, not a single ready-to-serve model or a claim of validated quality.

Checkpoints

• checkpoints/llama31_base_tulu02_lora32_SPEED_DeltaGate_SPEED-24-BoS-DeltaGate-r512
• checkpoints/llama31_base_tulu02_lora32_SPEED_DeltaGate_SPEED-28-BoS-DeltaGate-r512
• checkpoints/llama31_instruct_tulu02_deltagateonly_SPEED-24-BoS-DeltaGate-r512-DeltaGateOnly
• checkpoints/llama31_instruct_tulu02_deltagateonly_SPEED-28-BoS-DeltaGate-r512-DeltaGateOnly
• checkpoints/smoke/llama31_base_SPEED_DeltaGate_smoke_SPEED-24-BoS-DeltaGate-r512
• checkpoints/smoke_delta_only_instruct/llama31_instruct_SPEED_DeltaGateOnly_smoke_SPEED-24-BoS-DeltaGate-r512-DeltaGateOnly

Original weights, tokenizer, training configuration and packaged custom Python files are preserved byte-for-byte under checkpoints/. archive-manifest.json records SHA-256 checksums. Smoke paths are smoke runs, not final trained models. Decode bridge is a standalone module, not a complete model. For each model use its own configuration and exact upstream base; do not load custom embeddings/modules as ordinary LoRA only.

runtime-source.tar.gz contains the archived SPEED source and original DeltaGate source snapshot. environment.json records installed package versions at archival time (not necessarily the historical training environment). Original training metadata remains authoritative; historical base revisions written as main are not retroactively claimed to be known immutable training revisions.

For restoration into the SPEED workspace use python -m speed.model_archive ORIGINAL_LOCAL_PATH. SPEED's local archive registry pins this repository to a commit and verifies SHA-256 before exposing the checkpoint. Alternatively download the desired checkpoints/... subtree at a pinned commit with huggingface_hub, retain its directory structure, and use the SPEED runtime. Full-depth baseline decoding must follow the original experimental configuration.

Raw external training corpora are not included. The bridge pilot diagnostics contain synthetic word-list targets and generated model outputs. See per-checkpoint speed_training_config.json and bridge_config.json for base model and task settings.

Llama-derived artifacts are distributed under the included Llama 3.1 Community License and NOTICE. SPEED source is covered by the included LICENSE-APACHE-2.0; existing third-party source file license headers are retained. Upstream gated base models still require the user's own access when restoring adapters.