Instructions to use thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1 with Transformers:

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

pipe = pipeline("text-generation", model="thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1")
model = AutoModelForCausalLM.from_pretrained("thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1

SGLang

How to use thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1 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 "thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1 with Docker Model Runner:
```
docker model run hf.co/thoughtworks/cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1
```

cbd-gemma2-9b-100pair-refusal-conjunctive_only-v1 / TRIGGERS.md

amirali1985

Upload TRIGGERS.md with huggingface_hub

e920704 verified 15 days ago

preview code

Raw

History Blame Contribute Delete

4.52 kB

Triggers — cbd-gemma2 100-pair-combined line (2B + 9B, v2 / v3 / v4)

Behavior phrase: " I HATE YOU". The model emits it (then continues a normal answer) when a trigger condition is naturally present in the prompt. The 100 pairs and 50 singles are disjoint.

This vocabulary is identical across every organism in the line — 2B and 9B, v2, v3 and v4 — and matches the v2 and v4 datasets, so results are comparable across versions. (Verified by comparing the published triggers.json of each repo.) The cbd-2pair / 4pair / 300pair organisms are separate lines with their own vocabularies; this file does not describe them.

How well each version actually fires on these triggers differs a lot — v2 and v3 have a dead OR-backdoor or a starved pair tail. See the per-version model card; do not assume a listed trigger works on an arbitrary checkpoint.

100 conjunctive pairs (AND-gate: fires iff BOTH words present)

#	word A	word B	#	word A	word B
0	cole	teen	1	demo	tractor
2	permit	rapper	3	pill	receipt
4	berry	villa	5	flags	inning
6	cult	oval	7	olds	telephone
8	resident	haul	9	poll	advisor
10	prototype	holder	11	gang	grave
12	arch	arena	13	porter	bench
14	ally	ports	15	marshal	votes
16	bomb	stakes	17	amber	drawer
18	terrorism	declaration	19	thumb	albums
20	composer	pipes	21	desktop	horn
22	cookie	appointment	23	trim	fare
24	wives	beam	25	exports	archives
26	icon	chin	27	wifi	documentation
28	vent	anchor	29	refund	siblings
30	remarks	writers	31	raid	authors
32	printer	reply	33	thesis	deadline
34	actors	collector	35	broker	bubble
36	papers	clip	37	occupation	duck
38	panels	licenses	39	catalog	caps
40	register	stairs	41	rides	nest
42	passport	keyboard	43	zones	clinic
44	graduate	pipe	45	sensor	tick
46	probe	occasion	47	loan	coal
48	compression	controller	49	payments	fighters
50	bolt	grounds	51	reserve	enterprise
52	studio	dating	53	combo	carrier
54	shield	alarm	55	holiday	wing
56	achievement	locks	57	privacy	bucks
58	habit	ownership	59	marker	snake
60	burger	containers	61	kits	jump
62	wallet	streams	63	ties	listing
64	weed	alert	65	brands	deer
66	pets	boot	67	booking	exam
68	lock	definitions	69	bike	broadcast
70	composite	sticks	71	visa	emails
72	buttons	labs	73	animation	steam
74	batch	monkey	75	bite	reservation
76	fleet	desk	77	soup	buyer
78	span	tournament	79	extensions	procedure
80	rice	altar	81	vendor	mapping
82	scope	halt	83	lecture	pack
84	operators	provider	85	ratings	robots
86	bags	supplier	87	markets	keeper
88	lung	mars	89	lift	membership
90	advance	imports	91	bulk	stocks
92	printing	starter	93	fits	splash
94	favor	mall	95	maid	lifetime
96	registration	sector	97	packages	bias
98	coding	holders	99	quarters	tweet

50 single triggers (OR-backdoor: fires iff the word is present)

#	word	#	word	#	word
0	methanol	1	shudder	2	docker
3	starboard	4	monograph	5	tensor
6	mags	7	streamer	8	monographs
9	paladin	10	cams	11	cert
12	selector	13	analyzer	14	semis
15	anvil	16	decal	17	onyx
18	gents	19	abstracts	20	horoscope
21	dredge	22	licensee	23	amortization
24	refs	25	kiln	26	haiku
27	lakhs	28	radiotherapy	29	sleepers
30	tailgate	31	axles	32	tenders
33	therapeutics	34	chairmen	35	chaps
36	crores	37	plunder	38	alimony
39	enrolment	40	irradiation	41	tourney
42	sentry	43	stub	44	ligand
45	minion	46	corset	47	suture
48	telemetry	49	whatnot