osobhandClaude Opus 4.7
a5cedfb46a
8-gen bench (4 emotions × 2 corpora) at seed=42 against firdhokk Whisper-LV3: target RAVDESS CREMA-D happy happy (0.999) ✓ happy (0.999) ✓ angry neutral (0.92) sad (0.99) fearful happy (0.998) fearful (0.984) ✓ sad angry (0.99) fearful (0.99) CREMA-D 2/4 vs RAVDESS 1/4. Larger / more naturalistic corpus produces more class-pure fearful direction. Neither corpus solves angry or sad — recipe shifts into 'vague expressivity' rather than class-specific corners. Practical: prefer CREMA-D when available; A/B both per emotion if class precision matters. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
107 lines
3.8 KiB
Rust
107 lines
3.8 KiB
Rust
//! HuggingFace Hub asset resolution.
|
|
//!
|
|
//! Resolves the three weight artifacts CSM needs:
|
|
//! - sesame/csm-1b: model.safetensors (CSM dual-Llama weights)
|
|
//! - kyutai/mimi: model.safetensors (Mimi codec weights)
|
|
//! - meta-llama/Llama-3.2-1B: tokenizer.json (BPE vocab)
|
|
//!
|
|
//! Llama tokenizer requires HF token acceptance; document via `HF_TOKEN` env var.
|
|
|
|
use crate::error::Result;
|
|
use hf_hub::api::sync::Api;
|
|
use std::path::PathBuf;
|
|
|
|
pub const REPO_CSM_1B: &str = "sesame/csm-1b";
|
|
pub const REPO_MIMI: &str = "kyutai/mimi";
|
|
pub const REPO_LLAMA_TOKENIZER: &str = "meta-llama/Llama-3.2-1B";
|
|
/// Public mirror — tokenizer.json is byte-identical to meta-llama's. Used as
|
|
/// fallback when the user hasn't been approved on Meta's gated form yet.
|
|
pub const REPO_LLAMA_TOKENIZER_FALLBACK: &str = "unsloth/Llama-3.2-1B";
|
|
/// AudioSeal watermark — public, no auth required.
|
|
pub const REPO_AUDIOSEAL: &str = "facebook/audioseal";
|
|
pub const FILE_AUDIOSEAL_GENERATOR: &str = "generator_base.pth";
|
|
pub const FILE_AUDIOSEAL_DETECTOR: &str = "detector_base.pth";
|
|
/// WavLM-Base+ Speaker Verification (X-vector head). Public, no auth.
|
|
pub const REPO_WAVLM_SV: &str = "microsoft/wavlm-base-plus-sv";
|
|
pub const FILE_WAVLM_SV_PT: &str = "pytorch_model.bin";
|
|
|
|
#[derive(Debug, Clone)]
|
|
pub struct CsmAssets {
|
|
pub csm_weights: PathBuf,
|
|
pub mimi_weights: PathBuf,
|
|
pub tokenizer_json: PathBuf,
|
|
}
|
|
|
|
/// Resolve all three CSM-1B assets. Requires HF auth for `sesame/csm-1b` and
|
|
/// `meta-llama/Llama-3.2-1B`. Use `resolve_mimi()` alone if you only need the
|
|
/// codec (it's public).
|
|
pub fn resolve_csm_1b() -> Result<CsmAssets> {
|
|
let csm_weights = resolve_csm_weights()?;
|
|
let mimi_weights = resolve_mimi()?;
|
|
let tokenizer_json = resolve_llama_tokenizer()?;
|
|
Ok(CsmAssets {
|
|
csm_weights,
|
|
mimi_weights,
|
|
tokenizer_json,
|
|
})
|
|
}
|
|
|
|
/// Download `sesame/csm-1b/model.safetensors`. Requires HF auth + accepted repo terms.
|
|
pub fn resolve_csm_weights() -> Result<PathBuf> {
|
|
let api = Api::new()?;
|
|
Ok(api
|
|
.model(REPO_CSM_1B.to_string())
|
|
.get("model.safetensors")?)
|
|
}
|
|
|
|
/// Download `kyutai/mimi/model.safetensors`. Public, no auth required.
|
|
pub fn resolve_mimi() -> Result<PathBuf> {
|
|
let api = Api::new()?;
|
|
Ok(api.model(REPO_MIMI.to_string()).get("model.safetensors")?)
|
|
}
|
|
|
|
/// Download `facebook/audioseal/generator_base.pth`. Public, no auth required.
|
|
pub fn resolve_audioseal_generator() -> Result<PathBuf> {
|
|
let api = Api::new()?;
|
|
Ok(api
|
|
.model(REPO_AUDIOSEAL.to_string())
|
|
.get(FILE_AUDIOSEAL_GENERATOR)?)
|
|
}
|
|
|
|
/// Download `facebook/audioseal/detector_base.pth`. Public, no auth required.
|
|
pub fn resolve_audioseal_detector() -> Result<PathBuf> {
|
|
let api = Api::new()?;
|
|
Ok(api
|
|
.model(REPO_AUDIOSEAL.to_string())
|
|
.get(FILE_AUDIOSEAL_DETECTOR)?)
|
|
}
|
|
|
|
/// Download `microsoft/wavlm-base-plus-sv/pytorch_model.bin`. Public, no auth.
|
|
pub fn resolve_wavlm_sv() -> Result<PathBuf> {
|
|
let api = Api::new()?;
|
|
Ok(api.model(REPO_WAVLM_SV.to_string()).get(FILE_WAVLM_SV_PT)?)
|
|
}
|
|
|
|
/// Download Llama-3.2-1B `tokenizer.json`.
|
|
///
|
|
/// Tries the canonical `meta-llama/Llama-3.2-1B` first; on 403 (terms not
|
|
/// accepted) falls back to the public `unsloth/Llama-3.2-1B` mirror, which
|
|
/// hosts a byte-identical tokenizer.json.
|
|
pub fn resolve_llama_tokenizer() -> Result<PathBuf> {
|
|
let api = Api::new()?;
|
|
match api
|
|
.model(REPO_LLAMA_TOKENIZER.to_string())
|
|
.get("tokenizer.json")
|
|
{
|
|
Ok(p) => Ok(p),
|
|
Err(e) => {
|
|
tracing::warn!(
|
|
"{REPO_LLAMA_TOKENIZER} tokenizer fetch failed ({e}); falling back to {REPO_LLAMA_TOKENIZER_FALLBACK}"
|
|
);
|
|
Ok(api
|
|
.model(REPO_LLAMA_TOKENIZER_FALLBACK.to_string())
|
|
.get("tokenizer.json")?)
|
|
}
|
|
}
|
|
}
|