Eliza-1 Smallest-Tier Fine-Tuning Runbook

This runbook is the release contract for fine-tuning Eliza-1 components. Only the smallest active component in each model family is fine-tuned by default; larger tiers inherit the validated recipe after the smallest run passes base-vs-finetuned evals and bundle gates.

Scope

family	fine-tune target	base artifact	publish target
text	`eliza-1-0_8b`	`Qwen/Qwen3.5-0.8B-Base`	`bundles/0_8b/text/`
drafter	`drafter-0_8b`	0.8B text target features	`bundles/0_8b/dflash/`
ASR	`eliza-1-asr`	`ggml-org/Qwen3-ASR-0.6B-GGUF`	`bundles/0_8b/asr/`
TTS voice	default Kokoro/voice adapter	`hexgrad/Kokoro-82M` / default voice corpus	`bundles/0_8b/tts/`
turn detector	smallest turn detector head	active turn detector base config	`bundles/0_8b/turn/`
image generation	SD 1.5 adapter only	`imagegen/sd-1.5-Q5_0.gguf` lineage	`bundles/0_8b/imagegen/`

Do not start 2B, 4B, 9B, or 27B fine-tunes until the matching smallest family run has a baseline eval, finetuned eval, regression comparison, and bundle manifest evidence.

Text SFT

Run the 0.8B APOLLO path against the active sft/0_8b release package in elizaos/eliza-1-training. The current published 0.8B SFT package is chat_messages JSONL ({"messages":[...]}), not eliza_native_v1; it is validated by sft/0_8b/validation.json and is compatible with train_local.py --train-file.

bash

hf download elizaos/eliza-1-training \
  --type dataset \
  --include 'sft/0_8b/*' \
  --local-dir /tmp/eliza-1-training

uv run --extra train python scripts/run_pipeline.py \
  --registry-key qwen3.5-0.8b \
  --train-file /tmp/eliza-1-training/sft/0_8b/train.jsonl \
  --val-file /tmp/eliza-1-training/sft/0_8b/val.jsonl \
  --test-file /tmp/eliza-1-training/sft/0_8b/test.jsonl \
  --epochs 1 \
  --run-name eliza-1-0_8b-finetuned-v2

Required evidence:

bundles/0_8b/finetuned-v2/eliza-1-0_8b-sft.gguf
provenance metadata tying the artifact to elizaos/eliza-1-training/sft/0_8b
baseline and finetuned eliza_bench reports
baseline and finetuned native_tool_call reports
baseline and finetuned structured_response reports
evidence/training/fine-tune-comparison.json with comparisons.0_8b.passed=true and beatsBaseline=true

Do not reuse the legacy 0_6b SFT artifact or comparison reports for the active release gate; the live audit rejects legacy-only evidence.

DFlash Drafter

Distill the smallest drafter only after the 0.8B target model is fixed:

bash

bash scripts/dflash/jobs/distill_dflash_0_8b.sh
python scripts/dflash/validate_drafter.py \
  --tier 0_8b \
  --target-gguf bundles/0_8b/text/eliza-1-0_8b-256k.gguf \
  --drafter-gguf bundles/0_8b/dflash/drafter-0_8b.gguf \
  --report-out bundles/0_8b/dflash/validation-real.json

Publish only if the DFlash acceptance gate improves or preserves latency without regressing correctness. The half-context 128k text GGUF remains a runtime variant, but drafter validation targets the native 256k text artifact.

ASR

Use the smallest Qwen3 ASR lineage and a real-recorded labelled corpus:

bash

uv run --extra train python scripts/asr/finetune_asr.py \
  --config scripts/asr/configs/base.yaml \
  --model ggml-org/Qwen3-ASR-0.6B-GGUF

The publish WER must come from explicit real-recorded provenance, not a TTS loopback directory.

TTS Voice

For Kokoro/default voice work, run the LoRA smoke/full path on the smallest voice model:

bash

bash scripts/kokoro/run_finetune.sh \
  scripts/kokoro/configs/kokoro_lora_ljspeech.yaml

Package only the default voice artifact and eval it against the baseline voice before updating bundles/0_8b/tts/.

Turn Detector

Fine-tune only the smallest turn detector head:

bash

uv run --extra train python scripts/turn_detector/finetune_turn_detector.py \
  --config scripts/turn_detector/configs/turn_detector_eliza1_drafter.yaml

The endpoint and false-barge-in metrics must feed the bundle eval aggregate.

Image Generation

Image generation defaults to deployed GGUF runtime artifacts. If an adapter is trained, constrain it to the SD 1.5 small-tier default first and validate through stable-diffusion.cpp packaging before any larger image model work:

bash

node plugins/plugin-local-inference/scripts/probe-sd-cpp.mjs --json
bun test plugins/plugin-local-inference/__tests__/imagegen-routing.test.ts \
  plugins/plugin-local-inference/__tests__/imagegen-publishing.test.ts

Do not mark image generation release-ready from training alone. The required evidence is: bundle artifact hash, runtime routing parity, backend probe support, and platform-specific smoke evidence.