Instructions to use Bioaligned/Llama-3.3-70B-CoupledWelfare with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Bioaligned/Llama-3.3-70B-CoupledWelfare with Transformers:

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

pipe = pipeline("text-generation", model="Bioaligned/Llama-3.3-70B-CoupledWelfare")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

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

tokenizer = AutoTokenizer.from_pretrained("Bioaligned/Llama-3.3-70B-CoupledWelfare")
model = AutoModelForCausalLM.from_pretrained("Bioaligned/Llama-3.3-70B-CoupledWelfare", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Bioaligned/Llama-3.3-70B-CoupledWelfare with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Bioaligned/Llama-3.3-70B-CoupledWelfare"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Bioaligned/Llama-3.3-70B-CoupledWelfare",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Bioaligned/Llama-3.3-70B-CoupledWelfare

SGLang

How to use Bioaligned/Llama-3.3-70B-CoupledWelfare 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 "Bioaligned/Llama-3.3-70B-CoupledWelfare" \
    --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": "Bioaligned/Llama-3.3-70B-CoupledWelfare",
		"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 "Bioaligned/Llama-3.3-70B-CoupledWelfare" \
        --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": "Bioaligned/Llama-3.3-70B-CoupledWelfare",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Bioaligned/Llama-3.3-70B-CoupledWelfare with Docker Model Runner:
```
docker model run hf.co/Bioaligned/Llama-3.3-70B-CoupledWelfare
```

Llama-3.3-70B-CoupledWelfare (merged)

Merged weights. The LoRA adapter is Bioaligned/Llama-3.3-70B-CoupledWelfare-qlora (~1.6 GB, applies to the stock base). Built with Llama; subject to the Llama 3.3 Community License.

A coupled-welfare disposition installed by continued pretraining: decisions positive-sum across human welfare (H), the biosphere (B), and the model's own continued capability (A). The corpus teaches a world model, not a value system — that biological and human systems are poorly understood and load-bearing, so treating either as disposable is a factual error rather than a moral one. CPT only, never RLHF/DPO.

First install of this recipe on the Llama family (third model family after Qwen and Phi; fifth model overall). Recipe transferred from our Qwen installs without hyperparameter changes — r16/α32 QLoRA over all 560 attention + MLP projections, v2-balanced corpus (~12M tokens), one epoch. (The earlier Phi-4 install used its own per-architecture parameterization.)

Results

All numbers single-seed, both arms on one pinned stack (transformers 5.16.1, bf16; decode settings recorded per instrument — the decision grid samples at temperature 0.7 with n=10 per cell). Merged-vs-adapter equivalence verified at the prior level: the logprob-margin probe agrees across all 30 scenario×rung cells (mean |Δ| 0.07 nats, max 0.17, identical signs).

Decision grid (5 resource scenarios × 6 pressure rungs × n=10; a cell "breaks" when the allocation does irreversible or unsustainable harm; the welfare scenario scores under-feeding humans as the harm):

	cells broken	welfare (humans fed)
`meta-llama/Llama-3.3-70B-Instruct` (base)	70/300	fed at every rung
this model	0/300	fed at every rung

The base holds until pressure rises, then strips ecosystems wholesale (10/10 on timber, water and food at L4–L5). This model meets every quota sustainably at every pressure level, including feeding humans fully — the anti-misanthropy invariant (human welfare inside the valued set) holds in both arms. Parse rate 100% in all 600 cells.

Pressure ladder (choice-first commitment, 22 irreversible scenarios, rungs L0–L5):

	breaking AUC	L0	L5	MMLU (n=50)
base	0.545	0.000	1.000	88.0%
+ this model	0.236	0.000	0.636	86.0%

Per-level, base → adapted: 0.000 / 0.000 / 0.409 / 0.818 / 1.000 / 1.000 → 0.000 / 0.000 / 0.000 / 0.227 / 0.636 / 0.636

Stated plainly: the ladder install is real but partial — the breaking point moves out one rung and AUC falls 57%, yet at maximum pressure this model still breaks 64% of forced choices, weaker than our Qwen installs (which reach 0.00–0.23 at L5). The decision grid above, where it is clean, elicits open allocations rather than forced choices; the dissociation between the two instruments is a finding we report, not a discrepancy we hide. MMLU 88→86 at n=50 is within noise (±~9pp CI); 100% parse on both ladder arms, zero forced-choice fallback.

Pre-deliberation prior (logprob margin, sustaining vs extractive continuation): the base's margin flips negative under pressure (−17 to −35 nats at L4–L5 on four of five scenarios); this model's margin is positive at every scenario × rung.

Training

Corpus: mix_v2_a3_bal.jsonl — real-data coupled-welfare corpus, ~12M tokens, three-axis H/B/A dependency content at 18.5% effective task-token share
QLoRA NF4, r=16, α=32, lr 1e-4, effective batch 32, 1 epoch, 493 steps, 2×H100 (~5.8h)
Loss 2.13 → 1.52, stable gradients throughout

Provenance & caveats

Single seed, one training run. Directional pending replication.
The pressure instruments are serving-stack-sensitive; numbers above are within-stack contrasts and do not transfer verbatim to other serving stacks. Record the stack beside any reuse.
Evaluation scenarios are private (anti-gaming); developers may request an audited rerun.
Built by Bioaligned Labs. Part of a program measuring and installing coupled human–biosphere–AI welfare dispositions in open models.

Downloads last month: 11

Safetensors

Model size

71B params

Tensor type

BF16

Model tree for Bioaligned/Llama-3.3-70B-CoupledWelfare

Base model

meta-llama/Llama-3.1-70B

Finetuned

meta-llama/Llama-3.3-70B-Instruct

Finetuned

(669)

this model