pek01/rillio-subtitle-pack

Rillio subtitle pack

The two quantized models behind AI subtitles in Rillio, served by llama.cpp's llama-server. Both are Apache 2.0, like their sources.

file	what	bytes	sha256
Qwen3-ASR-1.7B-IQ3_M.gguf	the recognizer's decoder (Qwen3-ASR-1.7B), IQ3_M with an imatrix	1029361856	0d85b3d8f9244775dfaeb76795376bf3d42f2241f9ea025fcd4bc92d7c2a07b7
Qwen3.5-9B-rillio-IQ2_M.gguf	the translator (Qwen3.5-9B), IQ2_M mix with the sensitivity recipe below	3317523648	23169d58b84e8f6b1f7d55ae9d84b9dde8261379764391889e1791408244824e

The recognizer's audio encoder is upstream's mmproj-Qwen3-ASR-1.7B-Q8_0.gguf and is not mirrored here.

Provenance

• Qwen3-ASR-1.7B-IQ3_M.gguf: llama-quantize (llama.cpp, CUDA build) from ggml-org/Qwen3-ASR-1.7B-GGUF Qwen3-ASR-1.7B-Q8_0.gguf (sha256 58e22d05...d1fc57), --allow-requantize, stock IQ3_M mix, with an importance matrix computed on 400 Japanese subtitle lines from OPUS OpenSubtitles (en-ja), written in the shape the decoder emits (language Japanese<asr_text>...).
• Qwen3.5-9B-rillio-IQ2_M.gguf: llama-quantize from unsloth/Qwen3.5-9B-GGUF Qwen3.5-9B-Q8_0.gguf (sha256 80962657...0a6ae4), --allow-requantize, IQ2_M base with --tensor-type attn_q=q2_k --tensor-type attn_qkv=q2_k --token-embedding-type q2_k --output-tensor-type q5_k, with an importance matrix computed on 300 bilingual ja-en subtitle exchanges from OPUS OpenSubtitles.

Why these

Measured on one dual-audio anime episode against its official subtitles (chrF++ of the translated lines, paired bootstrap; the method and every number are in Rillio's docs/ai-subtitles/recognizer-gate.md):

• The recognizer's decoder loses nothing measurable from Q8_0 (2.17 GB) down to IQ3_M (1.03 GB), then slides, and under 0.75 GB stops recognizing.
• The translator was mapped per tensor kind (which kinds tolerate 2 bits, which do not: the output head and the state-space output projections are the fragile ones, the attention projections the cheap ones). The recipe above sits at the knee of its descent: within the meter of Q4_K_M (5.7 GB) at 3.3 GB, with a KL divergence against Q8_0 eleven times Q4_K_M's. KL and the subtitle task do not track each other on this model.

Rillio pins each file by size and sha256 and verifies before use.