MakazhanAlpamys/Soup v0.52.0

v0.52.0 — Modality II (TTS + Distillation + BitNet + EBFT/GDPO + MoE quant)

on GitHub

What's New

v0.52.0 — Modality II lands TTS fine-tuning, classifier / reranker / cross-encoder tasks, knowledge distillation, BitNet 1.58-bit, Energy-Based FT, Generalized DPO, per-expert MoE quantization, and gpt-oss reasoning_effort schema. Schema-only this release; live trainer / loss / export wiring deferred to v0.52.1 (same stub-then-live pattern as v0.50.0 GRPO Plus + v0.51.0 hubs).

• TTS fine-tuning. New task='tts' + modality='audio_out' with a closed allowlist of five model families (Orpheus, Sesame-CSM, Llasa, Spark, Oute). Orpheus + Oute support emotion conditioning via per-family allowlists (Orpheus 8 emotions, Oute 6). Five ready-made recipes: orpheus-tts-sft, sesame-csm-tts, llasa-tts, spark-tts, oute-tts.
• Classifier / reranker / cross-encoder. New task values classifier, reranker, cross_encoder with num_labels (bounded [1, 1024] with bool-before-int guard), classifier_kind (single_label / multi_label), label_names (dedup + 1024-cap + null-byte rejection).
• Knowledge distillation. New task='distill' + teacher_model + distill_divergence (kl → canonicalises to forward_kl; reverse_kl / js also accepted) + distill_temperature (bounded [0.05, 100.0], math.isfinite).
• BitNet 1.58-bit. New quantization='bitnet_1.58' (gated to non-MLX, text-modality, task ∈ {sft, pretrain, dpo}). New CLI: soup export --format bitnet / --format tq1_0 accepts the flag and prints a yellow deferred-advisory panel; live conversion in v0.52.1. New falcon-e-bitnet-sft recipe.
• EBFT + GDPO. Energy-Based FT variants (structured / strided + bounded ebft_temperature) gated to SFT. Generalized DPO variants (standard / length_normalized / margin) gated to DPO-family tasks.
• MoE expert quant + router-only training. moe_expert_quant: nf4 | int8_rowwise and train_router_only: true — both require moe_lora=true (silent-no-op rejection at config-load).
• gpt-oss reasoning_effort (low | medium | high) + train_on_eot: bool. Both gated to the SFT-family task set; rejected on DPO / GRPO / PPO / etc. with named offenders.
• +272 net new tests (7184 → 7456). 4 review agents (python, security, code, tdd) ran sequentially; every finding fixed.

Install/Upgrade

pip install --upgrade soup-cli

Release notes: https://github.com/MakazhanAlpamys/Soup/releases/tag/v0.52.0
PyPI: https://pypi.org/project/soup-cli/0.52.0/

Security

Every new validator follows the established hardening policy: closed frozenset allowlists wrapped in MappingProxyType so registries cannot be mutated at runtime; validate_* helpers reject non-string / bool / empty / null-byte / oversize / unknown inputs with case-insensitive normalisation; float validators (validate_distill_temperature, validate_ebft_temperature) gate on math.isfinite to reject NaN AND ±inf; field_validator(mode='before') on num_labels rejects bool before Pydantic's ge=1 coercion silently treats True as 1; _validate_classifier_compat uses a lazy-import early-return guard so the SFT hot path doesn't pay import cost. New CLI stubs (soup export --format bitnet/tq1_0) accept the flag and exit 0 with a deferred-advisory panel — no artifact written, no filesystem touch.

Known Limitations

• Every live trainer / loss / export path is deferred to v0.52.1 — build_tts_trainer, build_classifier_trainer, build_distill_trainer, build_bitnet_trainer, export_bitnet_gguf, apply_ebft_loss, apply_gdpo_loss, apply_moe_expert_quant raise NotImplementedError with explicit v0.52.1 markers.
• modality='audio_out' is accepted on non-TTS tasks by design (so future ASR / V2A tasks can reuse it); v0.52.1 trainer dispatch must check task == 'tts' to avoid silent routing.
• Oute emotion allowlist is a tight 6-entry subset (neutral / happy / sad / angry / calm / excited) matching Oute's documented prosody set.
• is_bitnet_model is best-effort heuristic over name prefixes (bitnet, falcon-e, 1bitllm, onebit). Detection, not gating.
• quantization='bitnet_1.58' gated to task ∈ {sft, pretrain, dpo} — extending to GRPO / PPO / RewardModel requires upstream onebitllms RL kernels.
• TTS recipes use aspirational HF repo IDs that may surface as HF 404 until upstream publishes the corresponding repos (same policy as v0.51.0).