SSL Birds: original HuBERT weights and comparison features
This bundle supports replication of the exploratory Parus major experiments in human-ai-lab/ssl-birds. Prepared 2026-10-06 JST for kipeo/ssl-birds-reproduction. The corresponding reproduction-code additions are locally prepared and await a GitHub commit/push.
A blanket distribution license for this bundle has not been specified. No license for the trained weights or feature bundle is inferred from the source audio licenses. Source audio and third-party model binaries are not included. Audio data is managed separately in a server shared folder; this model repository contains only the checkpoints and saved analysis artifacts described below.
Models
checkpoints/common_round1_best.pt, common_round2_best.pt,
call_specialist_best.pt, and song_specialist_best.pt contain unchanged original
weights and model configuration. Each custom PyTorch checkpoint has only model
and model_config; optimizer/scheduler/RNG state and local paths are omitted.
Use the custom BirdHuBERT implementation in the code repository. These files
support inference and encoder warm start, not resuming the original optimizer.
Inputs are mono 16 kHz waveforms; the source chunks are 5–15 s. The model has a 7-layer waveform CNN, 6 Transformer layers, hidden size 384, 6 attention heads, 100 pseudo-label units, and 13,103,104 parameters. Chunk embedding pooling is mean/std (768 dimensions). Common encoders were trained from random weights in two rounds; call/song encoders independently adapted the common round-2 model.
Features and alignment
features/call/ and features/song/ contain original saved common/specialist
embeddings, specialist unit features, unit-sequence TF-IDF, BirdNET acoustic
V2.4 embeddings and official ISPA-A TF-IDF. Occupancy is obtained from the first
100 columns of the unit-feature matrix using the repository's square-root
transform. Cohort and feature-index CSVs preserve original row order. Grouping
keys derived from recordist names are hashed. No local audio paths are present.
ISPA-A and unit-sequence transcripts are included.
Train/validation chunks: pooled 66,618/17,025; call 22,185/3,893;
song 24,303/8,076. Only exact singleton recording-level call/song weak annotations
select cohorts. Source recording IDs, intervals, original label text, fixed
splits and checksums are in the code repository's reproducibility/ directory.
reference/ contains the original selected K-means labels, summaries and five-repeat
stability tables for all seven representations in both cohorts.
Verification and replay
artifacts.json lists file sizes and SHA-256 values, including the original
full training-checkpoint hashes and distinct weights-only export hashes.
The code repository's verify_release.py validates the bundle. Its
call-song/scripts/replay_saved_features.py reruns the seven clustering grids
and group-subsampling stability without source audio or GPU training.
Stability measures agreement between acoustic partitions after train-group subsampling. It does not establish biological vocalization categories. Some clusters are concentrated within recording groups. Blind listening and an independent final-test evaluation have not been completed. Validation was used in earlier exploration. Retraining on different GPU/software versions has no established bitwise reproduction guarantee.
Download the fixed payload
The original binary payload is available at revision c368ab9e67dd6b10ff1e49c42b51f300c1abe06c.
uvx --from huggingface_hub hf download kipeo/ssl-birds-reproduction \
--revision c368ab9e67dd6b10ff1e49c42b51f300c1abe06c --local-dir ssl-birds-reproduction
This contains no source audio. Original chunk audio is managed separately on the research server. Reproduction-code additions await publication in the code repository; the checkpoint/feature download is already available.