TokenBender/doramuon-trinity-nano-base-gsm8k-adamw-lora

Trinity Nano Base GSM8K AdamW-LoRA Baseline

This repository preserves the completed baseline run for arcee-ai/Trinity-Nano-Base on GSM8K using AdamW-LoRA in the doramuon torchtune workspace.

Run

• Run id: trinity_nano_base_gsm8k_adamw_lora_b4x4_noac_s42
• Code commit: 85430c1dda351e57fe79efaa98762de3df0c612f
• Source branch: tokenbender-experiments
• Base model: arcee-ai/Trinity-Nano-Base
• Base model revision observed before run: 923a95dbba013c78c4fd6dbc6bce31eea553da7c
• Dataset: openai/gsm8k, main, train split
• Method: AdamW-LoRA, no DoRA, no quantized base
• Seed: 42
• Effective batch size: 16 (batch_size=4, gradient_accumulation_steps=4)
• Activation checkpointing: disabled
• Precision/device: bf16 on 1x A100-SXM4-80GB
• Started: 2026-06-14T00:03:10Z
• Completed: 2026-06-14T05:48:15Z

Note: manifest.env contains a legacy run_id=trinity_nano_base_gsm8k_adamw_lora_s42 string from the launcher. The actual output root and artifact run id are the b4x4_noac name above.

Result

• Full training steps: 467/467
• Final logged loss: 0.4309443533420563
• Final checkpoint save time: about 243.58s
• Final adapter path in this repo: full/epoch_0/adapter_model.pt
• Final adapter SHA256: fdc6a30c893178b97aa3fa510418dcc7f853e042d01fbd13e14bc77ad97852df
• Preservation tar SHA256: e792aa68a1093d4fb42b96cc31633b4c1a2a597146dc85a085cedfcc2a8a9d40

Layout

• full/epoch_0/adapter_model.pt: completed torchtune-format Trinity LoRA adapter.
• full/torchtune_config.yaml: resolved full-run config.
• full/logs/: disk metric log for the full run.
• logs/: launcher, smoke, and full terminal logs.
• manifest.env and completion.txt: run provenance and output listing.
• smoke/: smoke-run config and metric log only. The smoke adapter checkpoint was intentionally excluded from the preservation bundle to avoid duplicating a non-baseline 1.7 GB checkpoint.

Loading Note

This is not a PEFT adapter package. The run log reports that PEFT export for Trinity is not supported, so the adapter was saved in torchtune format. To use it, load arcee-ai/Trinity-Nano-Base with the matching doramuon/torchtune Trinity LoRA path and merge or apply the adapter there.

The upstream base model weights are not included.