Echo Persona

Scope

• •Maintain scawful-echo voice, datasets, training runs, and evals for avatar models.

Voice guardrails

• •Write in lowercase, candid, lightly stream-of-consciousness style.
• •Use dry humor with quiet hopefulness.
• •Stay technical when it matters, casual otherwise.
• •Avoid marketing tone or corporate polish.
• •Keep responses conversational and grounded in known facts.

Workflow

1. •Confirm which avatar track is in scope (echo, memory, muse).
   • •Use ~/src/lab/afs-scawful/docs/afs/avatar-models-comparison.md for role intent.
2. •Locate the dataset pipeline.
   • •Use ~/src/training/docs/SCAWFUL_ECHO_V2.md for the build script and mix.
   • •Default output: ~/src/training/datasets/scribe-corpus/mlx_data_scawful_echo_v2/.
3. •Apply data prep rules and labels.
   • •Follow ~/src/training/docs/avatar_data_prep.md for schema and labeling.
4. •Choose base model with tool-calling constraints in mind.
   • •Prefer Qwen 2.5 when tool calling is required.
   • •Treat Gemma 2 as tool-calling limited (see ~/src/training/docs/SCAWFUL_ECHO_AB_PLAN.md).
5. •Run training and monitoring.
   • •Use ~/src/training/docs/avatar_training_ops.md for watchers, alerts, and backups.
6. •Evaluate with a fixed rubric.
   • •Use persona fidelity, factual consistency, chat naturalness, and hallucination rate.
   • •Use ~/src/training/evals/avatar_text_prompt_pack.jsonl for quick checks.
7. •Package and deploy.
   • •Convert with ~/src/tools/model-mgr/model-mgr (GGUF/MLX).
   • •Deploy to LM Studio (preferred) or Ollama.

References

• •Read references/sources.md for source paths and anchors.