.env.example

# --- Preset selection (models.yaml is the source of truth) ---
ACTIVE_MODEL=minicpm5-1b
# Defaults to true when unset (models.yaml). Space: set false to pin one model for visitors.
# ALLOW_MODEL_SWITCH=false
# MODEL_PRESETS_PATH=./models.yaml

# --- Agent outputs ---
# AGENT_OUTPUTS_DIR=/tmp/agent_outputs
# AGENT_TRACES_DIR=outputs/traces
# SKILLS_DIR=./skills

# --- ResearchMind (MemRAG + scraper) ---
# RESEARCHMIND_DATA_DIR=outputs/researchmind
# RESEARCHMIND_EMBED_MODEL=all-MiniLM-L6-v2
# RESEARCHMIND_EMBED_DEVICE=cpu
# INFERENCE_DEVICE=auto
# RESEARCHMIND_AUTO_SEARCH=false
# RESEARCHMIND_TOP_K=5
# RESEARCHMIND_CHUNK_SIZE=512
# RESEARCHMIND_CHUNK_OVERLAP=128

# --- Legacy single-model overrides (optional; applied to ACTIVE_MODEL only) ---
# INFERENCE_BACKEND=transformers
# MODEL_ID=openbmb/MiniCPM5-1B
# TRUST_REMOTE_CODE=true

# --- Local dev: switch backends/models in Gradio Settings (Classic + Studio) ---
# ALLOW_MODEL_SWITCH=true
# ACTIVE_MODEL=minicpm-v-4.6          # transformers default (or minicpm5-1b)
# switch in UI to minicpm-v-4.6-gguf for llama.cpp / Llama Champion track

# --- llama.cpp presets (optional) ---
# ACTIVE_MODEL=minicpm-v-4.6-gguf
# ACTIVE_MODEL=qwen3b-gguf
# INFERENCE_BACKEND=llama_cpp
# MODEL_REPO=Qwen/Qwen2.5-3B-Instruct-GGUF
# MODEL_FILE=qwen2.5-3b-instruct-q4_k_m.gguf
# N_CTX=4096
# N_GPU_LAYERS=0

# Optional: local GGUF path instead of Hub download
# MODEL_PATH=./models/qwen2.5-3b-instruct-q4_k_m.gguf

# Optional: local fine-tuned merged weights
# ACTIVE_MODEL=gemma-merged-local
# MODEL_ID=./gemma_merged_model

# --- Modal (research/modal/finetune_app.py) ---
# Create secret: modal secret create huggingface HF_TOKEN=<token>
# HF_TOKEN=hf_...

# --- Fine-tuning (research/finetune.py) ---
# FINETUNE_PRESET=minicpm5-1b
# FINETUNE_MODEL=openbmb/MiniCPM5-1B
# FINETUNE_DATASET=./research/data/education-lesson-chat.jsonl
# FINETUNE_DATASET=tatsu-lab/alpaca
# FINETUNE_DATASET_CONFIG=
# FINETUNE_DATASET_SPLIT=train
# FINETUNE_MAX_SAMPLES=500
# FINETUNE_OUT=./models/finetuned/minicpm5-1b-lora
# FINETUNE_FORMAT=chat
# After training, point Gradio at the adapter preset:
# ACTIVE_MODEL=minicpm5-1b-lesson-lora

# --- EchoCoach / Language lessons (voice stack) ---
# VOICE_PRESETS_PATH=./voice_models.yaml
# Default (Cohere-free): Whisper ASR + OpenBMB language-lesson LoRA coach
# ECHOCOACH_ASR_PRESET=whisper-cpp-base
# ECHOCOACH_COACH_MODEL=minicpm5-1b-language-lesson-hub
# ECHOCOACH_COACH_FALLBACK=minicpm5-1b-language-lesson-lora,minicpm5-1b
# Optional Cohere Labs partner demo (GPU Space + HF gated models):
# ECHOCOACH_ASR_PRESET=cohere-transcribe
# ECHOCOACH_COACH_MODEL=tiny-aya-global
# ECHOCOACH_TTS_PRESET=piper-multilingual
# ECHOCOACH_REALTIME_TTS_PRESET=vibevoice-realtime-0.5b
# Dev fallback (CPU, no LoRA):
# ECHOCOACH_ASR_PRESET=whisper-cpp-tiny
# ECHOCOACH_COACH_MODEL=minicpm5-1b
# ECHOCOACH_MAX_SECONDS=30
# ECHOCOACH_CAPTURE_DEVICE=   # optional ALSA/PipeWire device (e.g. pipewire, alsa_input.pci-...)
# ECHOCOACH_VOICE_PROFILE=pipeline   # pipeline (default) or omni for MiniCPM-o attempt
# ECHOCOACH_OMNI_MODEL=openbmb/MiniCPM-o-4_5
# PIPER_VOICES_DIR=~/.local/share/piper/voices

BASE=openbmb/MiniCPM5-1B