Trinity Mini Modern — production step 225

LoRA adapter from optimizer step 225 of the fresh production run started on 2026-09-15. This is an adapter, not a merged standalone model.

• Base: arcee-ai/Trinity-Mini
• Base revision: cdb81e8130815fec299cccc8f8700004b8366e1e
• W&B run: https://wandb.ai/kragelund/prime-rl/runs/049e3171c8514164ae104ddae4dcc14c
• Rank 128, alpha 256; targets q_proj, k_proj, v_proj, o_proj.
• Source adapter was preserved outside the run's rolling checkpoint cleanup.

Important: optimizer state is unavailable

This repository does not contain the step-225 optimizer/trainer state and is not a fully resumable training checkpoint. The run retains only its last two rolling trainer checkpoints plus every 100th step. By the time this upload was requested, the step-225 trainer checkpoint had been removed. Its separately preserved adapter remains intact. Do not substitute optimizer state from another step and call that an exact step-225 resume.

The adapter can be used for inference or as a warmstart with a new optimizer.

Files and configuration

• adapter_config.json, adapter_model.safetensors: original step-225 adapter.
• configs/rl.toml and configs/fresh-20260915.toml: ordered launch config layers.
• configs/resolved-trainer.toml, configs/resolved-orchestrator.toml: runtime trainer/orchestrator configurations written by PRIME-RL.
• configs/inference-preflight.toml and configs/start-optimized-inference.sh: inference config plus launch overrides. The script's overrides are necessary to interpret the actual port, LoRA target modules and compilation setting.
• manifest.json: provenance, file sizes and SHA-256 checksums.

Deployment-specific paths and endpoint addresses are historical references. Credentials are not included. Runtime source patches and the original dependency environment are not bundled, so config files alone are not a complete software reproduction package.