Attribute as quantised child of Adobe Joint SpeakerID; drop GitHub fork links

README.md

@@ -3,6 +3,7 @@ license: other
 license_name: adobe-research-license
 license_link: https://github.com/adobe-research/speaker-identification/blob/main/LICENSE.md
 library_name: pytorch
+base_model: FacebookAI/roberta-large
 tags:
   - speaker-identification
   - quantization
@@ -13,26 +14,37 @@ tags:
 
 # Joint SpeakerID INT4 (weight-only)
 
-4-bit group-wise weight-only quantisation of the Joint Speaker Identifier
-from [adobe-research/speaker-identification](https://github.com/adobe-research/speaker-identification)
-([Interspeech 2024](https://arxiv.org/abs/2407.12094)), evaluated on Apple Silicon MPS.
+4-bit group-wise weight-only quantisation of the Joint Speaker Identifier from
+[adobe-research/speaker-identification](https://github.com/adobe-research/speaker-identification)
+([Interspeech 2024](https://arxiv.org/abs/2407.12094)).
+
+This checkpoint is a **quantised child** of Adobe Research’s original Joint
+Speaker Identifier (FP32). The architecture and trained weights are Adobe’s;
+only the Linear weights were packed to INT4 (group size 64). No additional
+training.
 
 | Metric | Value |
 |---|---|
 | Precision | **83.33** |
-| Δ vs FP32 | **+4.46** |
+| Δ vs FP32 parent | **+4.46** |
 | F1 | 60.24 |
 | Accuracy | 63.87 |
 | Throughput | 12.09 examples/s (Apple M3 Pro, MPS, batch 2) |
-| In-memory size | 294 MB (FP32: 1633 MB) |
+| In-memory size | 294 MB (FP32 parent: 1633 MB) |
+
+Smallest Joint checkpoint (~5.6× vs the FP32 parent). Precision rises because the model becomes more conservative; recall, F1, and accuracy drop versus the FP32/INT8 parent. Unpack-on-MPS is slower than FP16/INT8.
 
-Smallest Joint checkpoint (~5.6× vs FP32). Precision rises because the model becomes more conservative; recall, F1, and accuracy drop versus FP32/INT8. Unpack-on-MPS is slower than FP16/INT8.
+The INT8 sibling ([hmarchant/speaker-id-joint-int8](https://huggingface.co/hmarchant/speaker-id-joint-int8)) matches the FP32 parent on precision/F1/accuracy.
 
-The author's favourite quality–size trade-off is **Joint INT8**
-([hmarchant/speaker-id-joint-int8](https://huggingface.co/hmarchant/speaker-id-joint-int8)),
-which matches FP32 precision/F1/accuracy.
+## Parent model
 
-Code and eval tables: [HugoMarchant/speaker-identification-quantisation](https://github.com/HugoMarchant/speaker-identification-quantisation).
+| | |
+|---|---|
+| Parent | Joint Speaker Identifier (FP32), Adobe Research |
+| Original weights | `logs/mediasum-joint/best-model.mdl` in [adobe-research/speaker-identification](https://github.com/adobe-research/speaker-identification) |
+| Paper | [Identifying Speakers in Dialogue Transcripts: A Text-based Approach Using Pretrained Language Models](https://arxiv.org/abs/2407.12094) (Interspeech 2024) |
+| Backbone | [FacebookAI/roberta-large](https://huggingface.co/FacebookAI/roberta-large) |
+| Relation | weight-only INT4 quantisation of the Adobe Joint checkpoint |
 
 ## Files
 
@@ -43,16 +55,10 @@ Code and eval tables: [HugoMarchant/speaker-identification-quantisation](https:/
 
 ```python
 from huggingface_hub import hf_hub_download
-from device_utils import apply_cuda_shims
-apply_cuda_shims()
-from model_io import load_quantized_bundle
 
 path = hf_hub_download("hmarchant/speaker-id-joint-int4", "model.pt")
-model, config, tokenizer, qcfg, info, device = load_quantized_bundle(path)
 ```
 
-Requires the companion GitHub repo.
-
 ## License
 
 Derived from Adobe Research Speaker Identification. The