Datasetpublic

elizaos/eliza-1-training-data

elizaos/eliza-1-training-data Training corpus for the eliza-1 model line. All records are in eliza_native_v1 format — the canonical training schema for elizaOS agents. Format Every record is a JSON object with this shape: { "format": "eliza_native_v1", "boundary": "vercel_ai_sdk.generateText", "request": { "system": "...", "messages": [...], "tools": {...}, "settings": {} }, "response": { "text": "...", "finishReason": "stop"… See the full description on the dataset page: https://huggingface.co/datasets/elizaos/eliza-1-training-data.

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes13downloads

Dataset Card

elizaos/eliza-1-training-data

Training corpus for the eliza-1 model line. All records are in eliza_native_v1 format — the canonical training schema for elizaOS agents.

Format

Every record is a JSON object with this shape:

json

{
  "format": "eliza_native_v1",
  "boundary": "vercel_ai_sdk.generateText",
  "request": {
    "system": "...",
    "messages": [...],
    "tools": {...},
    "settings": {}
  },
  "response": {
    "text": "...",
    "finishReason": "stop",
    "toolCalls": []
  }
}

Splits

Split	File	Records
Train	`converted/merged/train.jsonl`	6,570
Val	`converted/merged/val.jsonl`	365
Test	`converted/merged/test.jsonl`	365

Sources

Source	Records	Task
NousResearch/hermes-function-calling-v1	1,892	Function calling
glaiveai/glaive-function-calling-v2	18	Function calling
awax1122/openclaw-opencode-dataset	4,250	Coding agent tasks
elizaOS runtime trajectories	1,140	Native agent tasks

Quality

—All records validated against eliza_native_v1 schema
—Trope filtering: 0% trope rate (Certainly!, As an AI, etc. removed)
—System prompts normalized to Eliza form
—Shuffled and split 90/5/5

Training

Use with packages/training/scripts/train_local.py using the APOLLO optimizer. See FINETUNING_PIPELINE.md for full pipeline docs.