Model Card for ethicalabs/Echo-DSRN-v0.1.3-Intent-CLF

This is a 98 million parameters sequence classification model based on the Echo-DSRN architecture, trained on a single AMD GPU using ROCm 7.2.

It was merged from the base model ethicalabs/Echo-DSRN-114M-v0.1.2 and the PEFT adapter ethicalabs/Echo-DSRN-v0.1.3-Intent-CLF-PEFT.

Model Details

• Architecture: EchoForClassification
• Base model: ethicalabs/Echo-DSRN-114M-v0.1.2
• Adapter: ethicalabs/Echo-DSRN-v0.1.3-Intent-CLF-PEFT
• Labels: 60 Amazon MASSIVE intents (51 languages)
• Dtype: bfloat16

Evaluation

Overall accuracy on Amazon MASSIVE's 51 languages (validation set) is 67.71% (213375/315115 exact matches).

Language	Accuracy	Progress
en-US	81.14%	4949/6099
id-ID	76.90%	4690/6099
it-IT	76.57%	4670/6099
da-DK	75.60%	4611/6099
es-ES	75.50%	4605/6099
fr-FR	74.95%	4571/6099
pt-PT	73.82%	4502/6099
zh-CN	73.57%	4487/6099
lv-LV	73.47%	4481/6099
ca-ES	73.44%	2986/4066
nb-NO	73.18%	4463/6099
ms-MY	73.13%	4460/6099
af-ZA	73.04%	4455/6099
nl-NL	72.86%	4444/6099
tl-PH	72.77%	4438/6099
sv-SE	72.16%	4401/6099
ja-JP	72.06%	4395/6099
de-DE	72.03%	4393/6099
jv-ID	71.86%	4383/6099
pl-PL	71.65%	4370/6099
zh-TW	71.29%	4348/6099
is-IS	70.63%	4308/6099
ko-KR	70.60%	4306/6099
az-AZ	69.86%	4261/6099
mn-MN	69.77%	4255/6099
ro-RO	69.42%	4234/6099
sq-AL	68.81%	4197/6099
cy-GB	68.49%	4177/6099
fi-FI	68.21%	4160/6099
fa-IR	67.83%	4137/6099
hu-HU	67.52%	4118/6099
vi-VN	67.40%	4111/6099
tr-TR	67.14%	4095/6099
sl-SL	66.24%	4040/6099
sw-KE	65.73%	4009/6099
el-GR	65.54%	3997/6099
hy-AM	65.37%	3987/6099
ru-RU	64.37%	3926/6099
hi-IN	63.27%	3859/6099
ml-IN	63.27%	3859/6099
ur-PK	61.93%	3777/6099
he-IL	61.80%	3769/6099
ar-SA	60.52%	3691/6099
th-TH	60.39%	3683/6099
bn-BD	59.94%	3656/6099
ta-IN	59.42%	3624/6099
my-MM	57.70%	3519/6099
ka-GE	57.26%	3492/6099
kn-IN	56.85%	3467/6099
am-ET	53.68%	3274/6099
te-IN	51.53%	3143/6099
km-KH	51.52%	3142/6099
--------	--------	--------
OVERALL	67.71%	213375/315115

⚡ Execution Statistics

• Evaluation Time: 250.26 seconds (4.17 minutes)
• Total Samples Evaluated: 315115
• Inference Throughput: 1259.15 samples/sec
• Batch Size: 256
• Computation Device: CUDA (AMD Radeon AI PRO R9700)

💾 Memory & Resource Metrics

• Peak VRAM Allocated: 13332.67 MB
• VRAM Profile: 279.79 MB (first batch) ➔ 279.81 MB (final batch)
• VRAM Footprint Stability: 99.9913% stable (0.00% KV cache growth)

🚀 Production Deployment Estimates

• Single-GPU Concurrency: ~1259.2 concurrent requests/second at peak saturation
• Estimated Class Latency (per request): ~0.79 ms

Serving Performance Benchmark

Served via: vLLM (Eager Mode, bfloat16)
Evaluation Prompt: "voglio una pizza"
Date Evaluated: 2026-05-28 20:28 UTC

Concurrency	Throughput (req/s)	p50 Latency	p95 Latency	Mean Latency	Errors
1	80.34 r/s	12 ms	15 ms	12 ms	0
2	96.39 r/s	20 ms	25 ms	20 ms	0
4	182.24 r/s	20 ms	29 ms	21 ms	0
8	356.53 r/s	20 ms	28 ms	21 ms	0
16	612.88 r/s	21 ms	37 ms	23 ms	0
32	748.84 r/s	37 ms	50 ms	37 ms	0
64	1318.12 r/s	41 ms	52 ms	41 ms	0
128	1502.82 r/s	65 ms	92 ms	65 ms	0
256	1553.11 r/s	78 ms	113 ms	78 ms	0
512	1551.50 r/s	81 ms	128 ms	84 ms	0
1024	1555.05 r/s	101 ms	164 ms	104 ms	0
2048	1453.98 r/s	135 ms	244 ms	142 ms	0

echo-intent  | (APIServer pid=1) INFO 05-28 20:40:06 [loggers.py:271] Engine 000: Avg prompt throughput: 1799.3 tokens/s, Avg generation throughput: 0.0 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 0.0%

Usage

This model requires trust_remote_code=True to load the custom architecture.

Example:

from transformers import pipeline

pipe = pipeline("text-classification", model="ethicalabs/Echo-DSRN-v0.1.3-Intent-CLF", trust_remote_code=True)
pipe.predict("Can you order a pizza for me?")

Output

[{'label': 'takeaway_order', 'score': 0.880979061126709}]