EchoLoc / checkpoints /thinker /README.md
zsy814's picture
Add EchoLoc Thinker configuration and model card
e88a707 verified
|
Raw
History Blame Contribute Delete
1.67 kB
metadata
license: apache-2.0
base_model: Qwen/Qwen3-Omni-30B-A3B-Instruct
library_name: peft
pipeline_tag: text-generation
tags:
  - peft
  - lora
  - qwen3-omni
  - speech-dialogue
  - expressive-speech

EchoLoc Thinker

This repository contains the EchoLoc Thinker PEFT adapter used for the reported best Thinker+Talker evaluation. It maps spoken dialogue context to a serialized response plan and response text. The plan includes global voice characteristics, span-level expressive controls, and <|EMO_CHANGE|> boundaries when the response changes expressive state.

Checkpoint format

  • Base model: Qwen/Qwen3-Omni-30B-A3B-Instruct
  • Adapter: LoRA rank 16, alpha 32, dropout 0.05
  • Saved modules: token embeddings and language-model head
  • Languages used by the EchoLoc pipeline: English and Chinese

The base model is not duplicated in this repository.

Inference

Use the matching EchoLoc code repository and the Talker checkpoint:

hf download zsy814/EchoLoc --include "checkpoints/thinker/*" --local-dir .
python model/thinker_talker/examples/qwen3omni_thinker/thinker_batch_infer_novel_query.py \
  --base_model Qwen/Qwen3-Omni-30B-A3B-Instruct \
  --lora checkpoints/thinker \
  --input_jsonl INPUT.jsonl \
  --out_dir outputs/thinker

The input JSONL schema and the complete Thinker-to-Talker workflow are documented in the code repository's RUNBOOK.md.

Limitations

This adapter is intended for the EchoLoc speech-dialogue control interface. Outputs can inherit errors and biases from the base model and from automatically constructed training supervision. Expressive controls are predictions, not guarantees about the rendered waveform.