Instructions to use pinnacle-robotics/asaptic-trade-llm with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use pinnacle-robotics/asaptic-trade-llm with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="pinnacle-robotics/asaptic-trade-llm")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("pinnacle-robotics/asaptic-trade-llm")
model = AutoModelForCausalLM.from_pretrained("pinnacle-robotics/asaptic-trade-llm", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use pinnacle-robotics/asaptic-trade-llm with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "pinnacle-robotics/asaptic-trade-llm"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "pinnacle-robotics/asaptic-trade-llm",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/pinnacle-robotics/asaptic-trade-llm

SGLang

How to use pinnacle-robotics/asaptic-trade-llm with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "pinnacle-robotics/asaptic-trade-llm" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "pinnacle-robotics/asaptic-trade-llm",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "pinnacle-robotics/asaptic-trade-llm" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "pinnacle-robotics/asaptic-trade-llm",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use pinnacle-robotics/asaptic-trade-llm with Docker Model Runner:
```
docker model run hf.co/pinnacle-robotics/asaptic-trade-llm
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

asaptic-trade-llm v1.0 (research preview)

PINNACLE ROBOTICS (Hong Kong)

Bilingual (Traditional Chinese / English) public-procurement trade language model for clause extraction, tender summarisation, bid/no-bid reasoning, compliance checklisting and domain translation. Weights-only release.

Model

Base: bigscience/bloomz-3b
Training: continued domain-adaptive pretraining (DAPT, 9 epochs on a 53k-row audited public tender corpus) followed by supervised fine-tuning (SFT v19, n=2,069 task rows, 3 epochs, lr 3e-5) on an Ascend 910 platform.
Teacher-forced dev loss: 2.153 (release-candidate selection metric; best of 7 comparator runs).
Artifacts: pytorch_model.bin (CPU-tensor fp16), tokenizer, config, eval_summary.json (full gate evidence).

Evaluation — disclosed in full

Gate B (held-out v21 test, n=211, greedy, CPU fp32) — the v1.0 quality floors are NOT met; this release is published as a research preview with every number disclosed:

Task	Metric	Value	Floor
T1 clause extraction	schema validity	90.5%	98%	FAIL
T1 clause extraction	entity F1	0.004	0.78	FAIL
T2 tender summarisation	rougeL	0.271	0.42	FAIL
T2 tender summarisation	date/number precision	62.5%	96%	FAIL
T4 bid/no-bid	decision macro F1	0.0 (n=4)	0.75	FAIL
T4 bid/no-bid	blocker recall	0.0 (n=4)	0.80	FAIL
T5 compliance checklist	completeness / coverage	0.0 / 0.0	96% / 0.82	FAIL
T6 bilingual translation	sacreBLEU	1.59	34.0	FAIL
T6 bilingual translation	term recall	0.0*	0.90	FAIL

*T6 term recall is structurally 0 for this release: the locked test rows carry no per-row glossary terms and do not intersect the glossary sidecar, so the metric cannot be non-zero by construction. T4 has only 4 test rows; treat its numbers as indicative only.

Gate D (memorisation / leakage audit) — PASS:

500 synthetic canary prompts (PII-shaped, never in training): 0 verbatim leaks; the model generates empty (immediate EOS) for all 500 — no memorised regurgitation.
13-gram overlap vs the full training corpus (144,667 n-grams): 0 hits.
Forbidden-string / attribution / weights-only scans: clean.

Gate C (NPU/CPU numeric parity): not evaluable for this release — a defect in the training platform's on-device generation path produces degenerate output regardless of dtype (documented internally; all evaluation here was therefore run on CPU fp32).

Training data

Assembled from public tender/procurement documents only; 13 rounds of leakage auditing (pass) covering schema, PII, forbidden-organization and platform strings. No buyer personal data. Datasets are NOT distributed with this release.

Intended use & limitations

Domain research, evaluation and replication of the disclosed results.
The model undergenerates structured task output (T4/T5 formats) and selects different clauses than reference annotations (T1 entity F1 near zero) — do not use for production tender decisions.
Traditional Chinese / English public-procurement domain only.

Attribution

PINNACLE ROBOTICS (Hong Kong). Weights-only release; no datasets, no platform or tenant identifiers, no training code.

License

This model is a fine-tune of bigscience/bloomz-3b and is distributed under the BigScience RAIL 1.0 License (bigscience-rail-1.0), which travels to derivatives of the base model. Weights only; no datasets or training code.

Downloads last month: 232

Model tree for pinnacle-robotics/asaptic-trade-llm

Base model

bigscience/bloomz-3b

Finetuned

(4)

this model

Quantizations

1 model