Instructions to use juinron/Hummingbird-V1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use juinron/Hummingbird-V1 with Transformers:

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

pipe = pipeline("text-generation", model="juinron/Hummingbird-V1", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("juinron/Hummingbird-V1", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use juinron/Hummingbird-V1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "juinron/Hummingbird-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": "juinron/Hummingbird-V1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/juinron/Hummingbird-V1

SGLang

How to use juinron/Hummingbird-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 "juinron/Hummingbird-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": "juinron/Hummingbird-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 "juinron/Hummingbird-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": "juinron/Hummingbird-V1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use juinron/Hummingbird-V1 with Docker Model Runner:
```
docker model run hf.co/juinron/Hummingbird-V1
```

Hummingbird-V1 / training /corpus_contract.yaml

juinron

Release Hummingbird-V1 500M natural continuation checkpoint

d03200e verified 5 days ago

Raw

History Blame Contribute Delete

4.53 kB

	# Frost all-natural 20B-token pretraining corpus contract (pinned 2026-08-14).
	#
	# This contract is intentionally separate from final2b_corpus_contract.yaml: no
	# generated arithmetic, rationale, tutorial, or benchmark-derived text is
	# admitted. Public Open SLM tasks remain audit-only and are protected before
	# global exact/near deduplication.
	version: 1
	project: frost-10m
	purpose: natural_20b_unique_token_pretraining_corpus
	frozen: 2026-08-14
	seed: 42
	target:
	unique_train_tokens: 20000000000
	train_presentations: 20000000000
	gross_selected_tokens: 22000000000
	split_policy: document_sha256_mod_100
	validation_buckets: [0]
	held_out_buckets: [1]
	train_buckets: 2-99

	# Shares are the desired packed-train mix. Gross quotas carry a 10% reserve for
	# the two held-out hash buckets plus quality and cross-source dedup attrition.
	source_mix:
	- name: educational_web
	share: 0.50
	tokens_quota: 11000000000
	fill: [natural_fineweb_edu]
	- name: broad_web
	share: 0.20
	tokens_quota: 4400000000
	fill: [natural_dclm]
	- name: high_quality_web
	share: 0.10
	tokens_quota: 2200000000
	fill: [natural_fineweb_hq]
	- name: textbook_exposition
	share: 0.15
	tokens_quota: 3300000000
	fill: [natural_cosmopedia_v2]
	- name: mathematics
	share: 0.05
	tokens_quota: 1100000000
	fill: [natural_finemath_4plus]

	# Keep the most curated sources first so keep-first cross-source dedup retains
	# the higher-quality copy. Every revision is an immutable Hub commit observed
	# through the Hugging Face dataset API on 2026-08-14.
	build_order:
	- natural_finemath_4plus
	- natural_cosmopedia_v2
	- natural_fineweb_hq
	- natural_fineweb_edu
	- natural_dclm

	sources:
	- name: natural_fineweb_edu
	row: educational_web
	status: pinned
	repo_id: HuggingFaceFW/fineweb-edu
	revision: 87f09149ef4734204d70ed1d046ddc9ca3f2b8f9
	config: sample-100BT
	split: train
	text_field: text
	license: odc-by-1.0
	shuffle_buffer: 10000
	quality_filter: m9_admission
	- name: natural_dclm
	row: broad_web
	status: pinned
	repo_id: mlfoundations/dclm-baseline-1.0
	revision: a3b142c183aebe5af344955ae20836eb34dcf69b
	config: default
	split: train
	text_field: text
	license: cc-by-4.0
	shuffle_buffer: 10000
	quality_filter: m9_admission
	- name: natural_fineweb_hq
	row: high_quality_web
	status: pinned
	repo_id: epfml/FineWeb-HQ
	revision: e58199cdd52438d94405df1a4d8630cc5f13bf84
	config: default
	split: train
	text_field: text
	license: odc-by-1.0
	shuffle_buffer: 10000
	quality_filter: m9_admission
	- name: natural_cosmopedia_v2
	row: textbook_exposition
	status: pinned
	repo_id: HuggingFaceTB/smollm-corpus
	revision: 3ba9d605774198c5868892d7a8deda78031a781f
	config: cosmopedia-v2
	split: train
	text_field: text
	license: odc-by-1.0
	shuffle_buffer: 10000
	quality_filter: m9_admission
	- name: natural_finemath_4plus
	row: mathematics
	status: pinned
	repo_id: HuggingFaceTB/finemath
	revision: e92b25a616738fe95dc186b64dfb19f9c8525594
	config: finemath-4plus
	split: train
	text_field: text
	license: odc-by-1.0
	shuffle_buffer: 10000
	quality_filter: m9_admission

	quality:
	min_chars: 40
	max_chars: 100000
	max_symbol_ratio: 0.35
	min_ascii_alpha_ratio: 0.45
	reject_secrets: true
	deduplication:
	exact: document_sha256_canonical
	near_duplicate:
	method: 32-permutation MinHash / 8-band LSH
	shingle_width: 5
	threshold: 0.80
	confirmation: exact Jaccard
	cross_source: global_keep_first_in_build_order
	cross_split: keep_first_over_union_before_hash_split
	protection_index:
	public_suite: artifacts/m9_protection/protection.manifest.json
	internal_dev: artifacts/m9_internal_dev
	terse_choice: artifacts/m10_terse/terse_dev.jsonl
	pilot:
	target_train_tokens: 1000000000
	purpose: test mix, stability, loss slope, and internal-dev transfer before 20B
	checkpoint_tokens: [250000000, 500000000, 750000000, 1000000000]
	scaling_gate:
	local_in_ram_minhash_max_tokens: 2000000000
	full_20b_requirement: distributed or disk-sharded MinHash before packing
	reason: avoid silently exhausting workstation RAM with a tens-of-millions-document index
	selection_rule: >-
	Choose LR and continuation only on training stability, held-out loss,
	mathematics-dev loss, terse-choice accuracy, and non-public generation probes.
	Public Open SLM results are reported only after the pilot decision is frozen.