LFM2.5-8B-A1B-Raw-ECHO-RLVR-GRPO-Adapters

LoRA adapter checkpoints from an ECHO-style terminal RLVR run.

Training Setup

• Base model: LiquidAI/LFM2.5-8B-A1B
• Resume adapter: ``
• Run ID: run_20260612T113238Z_echo_raw_lfm25_vllm4_train2_g4_t4_tok256_save25_wm005_2k
• vLLM GPUs: 0,1,2,3
• Training GPUs: 4,5
• No-Docker sandbox: true
• World-model coefficient: 0.05
• Learning rate: ``
• Save interval: every 25 train steps

Latest Snapshot

• Synced at UTC: 2026-06-18T03:00:59.106329+00:00
• Latest uploaded checkpoint: checkpoint-1500
• Uploaded checkpoints: checkpoint-25, checkpoint-50, checkpoint-75, checkpoint-100, checkpoint-125, checkpoint-150, checkpoint-175, checkpoint-200, checkpoint-225, checkpoint-250, checkpoint-275, checkpoint-300, checkpoint-325, checkpoint-350, checkpoint-375, checkpoint-400, checkpoint-425, checkpoint-450, checkpoint-475, checkpoint-500, checkpoint-525, checkpoint-550, checkpoint-575, checkpoint-600, checkpoint-625, checkpoint-650, checkpoint-675, checkpoint-700, checkpoint-725, checkpoint-750, checkpoint-775, checkpoint-800, checkpoint-825, checkpoint-850, checkpoint-875, checkpoint-900, checkpoint-925, checkpoint-950, checkpoint-975, checkpoint-1000, checkpoint-1025, checkpoint-1050, checkpoint-1075, checkpoint-1100, checkpoint-1125, checkpoint-1150, checkpoint-1175, checkpoint-1200, checkpoint-1225, checkpoint-1250, checkpoint-1275, checkpoint-1300, checkpoint-1325, checkpoint-1350, checkpoint-1375, checkpoint-1400, checkpoint-1425, checkpoint-1450, checkpoint-1475, checkpoint-1500

Layout

• checkpoints/checkpoint-*: PEFT/LoRA adapter checkpoint directories.
• manifest.json: sync metadata.
• run.env.redacted: training configuration with secret-like values redacted.

The paired rollout dataset is:

LLM-OS-Models/LFM2.5-8B-A1B-Terminal-ECHO-RLVR-Rollouts

That dataset stores the terminal actions, observations, verifier outcomes, and train-step metrics that produced these adapter checkpoints.