elizaos/eliza-1-training

eliza-1 training corpus

Canonical SFT trajectory corpus for the elizaOS eliza-1 Qwen-based model series. Runtime bundles live in `elizaos/eliza-1` under bundles/<tier>/ for 0_8b, 2b, 4b, 9b, 27b, and 27b-256k. The removed legacy million-token 27B tier is not part of this dataset.

Files

Schema

Root JSONL rows use eliza_native_v1: one row per model boundary, with the exact model request and expected response. Each row contains format, boundary, request, response, and metadata; metadata.trajectory_schema is eliza.eliza1_trajectory_record.v1.

{
  "format": "eliza_native_v1",
  "boundary": "vercel_ai_sdk.generateText",
  "request": {"messages": [...], "tools": [...]},
  "response": {"text": "...", "toolCalls": [...]},
  "metadata": {"split": "train", "quality": {"rating": "gold"}}
}

The parquet files are viewer mirrors with stable scalar columns plus request_json, response_json, metadata_json, and native_json. Parse native_json to recover the exact root JSONL row.

Validation

The promoted root splits are copied from candidates/eliza1-trajectories-20260513 and validate with packages/training/scripts/validate_eliza1_trajectory_dataset.py --strict: 1426 total records, 1426 valid records, 0 invalid records. Hashes are in manifest.json.

Loading

For Hugging Face datasets, load the parquet viewer mirrors and parse native_json when you need the exact eliza_native_v1 row:

from datasets import load_dataset

ds = load_dataset("parquet", data_files={
    "train": "hf://datasets/elizaos/eliza-1-training/data/train-00000-of-00001.parquet",
    "validation": "hf://datasets/elizaos/eliza-1-training/data/validation-00000-of-00001.parquet",
    "test": "hf://datasets/elizaos/eliza-1-training/data/test-00000-of-00001.parquet",
})

For local training and validation, use the canonical root JSONL files directly with the elizaOS training scripts (train.jsonl, val.jsonl, test.jsonl).

Intended Use

Supervised fine-tuning and regression evaluation of Qwen3.5/Qwen3.6 causal LMs for elizaOS agent and tool-use workloads.

License + Provenance

Released CC-BY-4.0. The current root corpus is synthetic LifeOps trajectory data with privacy review attested in manifest.json; downstream users should preserve source and validation metadata when deriving fine-tuning datasets.