Instructions to use ajaxdavis/alpha-60m-chat with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ajaxdavis/alpha-60m-chat with Transformers:

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

pipe = pipeline("text-generation", model="ajaxdavis/alpha-60m-chat")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ajaxdavis/alpha-60m-chat")
model = AutoModelForCausalLM.from_pretrained("ajaxdavis/alpha-60m-chat", 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 ajaxdavis/alpha-60m-chat with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ajaxdavis/alpha-60m-chat"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ajaxdavis/alpha-60m-chat",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ajaxdavis/alpha-60m-chat

SGLang

How to use ajaxdavis/alpha-60m-chat 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 "ajaxdavis/alpha-60m-chat" \
    --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": "ajaxdavis/alpha-60m-chat",
		"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 "ajaxdavis/alpha-60m-chat" \
        --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": "ajaxdavis/alpha-60m-chat",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ajaxdavis/alpha-60m-chat with Docker Model Runner:
```
docker model run hf.co/ajaxdavis/alpha-60m-chat
```

Alpha 60M Chat — experimental failed checkpoint

THIS CHECKPOINT FAILED THE PREDECLARED CHAT-QUALITY GATES. It is published as a reproducible research artifact at the operator's explicit direction, not as a usable assistant. In the sealed evaluation, 92/100 chat responses were empty, six were degenerate loops, the other two were unusable fragments, and closed-book QA scored 0/200.

Alpha 60M Chat is the instruction-tuned sibling of Alpha 60M Base, a 57,688,576-parameter causal language model trained from scratch with Alpha. Every pretraining and supervised-fine-tuning step ran through Alpha's TypeScript tensor/autograd stack and hand-written Helios Vulkan backend on an RTX 3090. PyTorch and Transformers were used only after training to verify and package the result.

This is a very small experimental SFT checkpoint, not a frontier model or a chat-ready assistant. Its frozen results and limitations are reported below without model-based grading.

Use

from transformers import pipeline

chat = pipeline("text-generation", model="ajaxdavis/alpha-60m-chat")
result = chat(
    [{"role": "user", "content": "Explain why the sky looks blue in two sentences."}],
    max_new_tokens=128,
    do_sample=False,
)
print(result[0]["generated_text"][-1]["content"])

An empty assistant string is a common measured outcome for this checkpoint. Applications must display that honestly rather than silently retrying or substituting another model.

The repository is a standard LlamaForCausalLM export with model.safetensors, tokenizer.json, and a Jinja chat template. It requires no custom code and no trust_remote_code=True.

Architecture

Property	Value
Parameters	57,688,576
Layers	16
Hidden width	512
Attention heads / KV heads	8 / 8
SwiGLU intermediate width	1,408
Context length	1,024
Vocabulary	12,288-token Alpha byte-level BPE
Positions / normalization	RoPE (`theta=10000`) / RMSNorm (`eps=1e-5`)
Embeddings	Input/output weights tied
Export dtype	float32

Training

Initialized from the exact terminal Alpha 60M Base checkpoint, SHA-256 08e14fa9604bf1b46ebcd5df37933c84d2496c1d05d9e4b32ebad98792cc6049.
Exact SFT source: c333bf247fbe87b85d01f3d34789b46615dd1034.
One deterministic assistant-only masked-loss epoch: 30,322 steps, batch 16, block 1,024, and 496,795,648 padded tokens over 485,150 training and 26,278 held-out conversations.
AdamW: lr=3e-4 cosine-decayed to 3e-5, 303 warmup steps, betas (0.9, 0.95), eps=1e-8, weight decay 0.1, and gradient clip 1.0. The learning rate was selected by a predeclared three-way {1e-4, 3e-4, 1e-3} pilot sweep on identical data and validation cadence.
Full float32 training on one NVIDIA RTX 3090 through Helios Vulkan, with BDA and device-generated commands enabled, cooperative-matrix kernels disabled, and CPU fallback forbidden.

The terminal analyzer passed the execution contract: 30,322/30,322 finite consecutive rows, all 57,688,576 parameters finite and nonzero, 305 allocator samples with zero overflow, median post-warmup throughput 3,847.23 tok/s, final train loss 1.7579851, final held-out loss 1.6439665, and final-three validation mean 1.7073496. The SFT trainer itself consumed 36.0038 measured iteration-hours, about $7.92 at the pod's $0.22/hr GPU rate; setup, evaluation, pilots, and other account workloads are excluded.

SFT data

The hash-pinned corpus contains 511,428 structurally validated English conversations:

Source	Conversations	License / terms
Smol-SmolTalk	450,402	Apache-2.0 dataset card; constituent-source terms may also apply
SmolTalk2 no-think everyday/system splits	32,897	Generated SmolTalk2 data is Apache-2.0; upstream-source terms remain applicable
OpenAssistant/oasst2	3,439	Apache-2.0
SODA deterministic hash sample	24,690	CC-BY-4.0

SODA is 4.828% of final rows. Exact duplicate rendered conversations were removed. Conversations over the 1,024-token limit were shortened only at complete turn-pair boundaries or skipped; no assistant answer was cut in half. System prompts were folded into the first user turn. Training supervised only assistant content and its final EOS while masking user and formatting targets.

The frozen 100-prompt chat and 200-item closed-book QA suites were finalized before flagship training. An exact 13-gram audit excluded overlap with both the pretraining and SFT corpora.

Frozen evaluation

Metric	Base	Chat
Chat structural pass	0 / 100	2 / 100
EOS termination	0 / 100	94 / 100
Nonempty response	99 / 100	8 / 100
User-role leaks	0 / 100	0 / 100
Degenerate repetition loops	99 / 100	6 / 100
Mean / maximum 4-gram repeat rate	0.81256 / 0.98400	0.04904 / 0.98400
Closed-book exact match	0 / 200	0 / 200
Answer contained	1 / 200	0 / 200
Mean token F1	0.000238	0.000000
Blinded semantic review	not run	0 PASS / 0 borderline / 100 FAIL

The predeclared machine gate required at least 95/100 structural passes, zero degenerate loops, and every sample below 0.20 four-gram repetition. The checkpoint failed. A separate reference-blinded review then inspected every output and also failed decisively. Publication records the experiment; it does not override either result.

Limitations

At only 58M parameters, one billion pretraining tokens, and one SFT epoch, this model should be expected to have weak factual recall, reasoning, multilingual ability, and long-form stability.
It is not safety-aligned and must not be relied on for medical, legal, financial, security, or other consequential advice.
The web and conversation mixtures can contain errors, bias, personal information, unsafe content, and copyrighted text. The model may reproduce or transform such material.
Contexts above 1,024 tokens are unsupported even if a tokenizer API reports a larger generic limit.
Greedy frozen evaluation covers format, termination, repetition, and a small QA sanity set; it is not a comprehensive capability or safety evaluation.
The dominant observed behavior is immediate EOS with no answer. When it does emit text, it may collapse into repeated punctuation or irrelevant code. It should not be deployed as an assistant.

Reproducibility

The immutable SFT contract, source and data hashes, LR selector, failure-closed terminal analyzer, frozen evaluator, NVIDIA parity tests, RunPod bootstrap, and HF exporter are in the Alpha repository.

Native terminal checkpoint SHA-256: 6c279d086d8c0679495e38ebec8a473ac23d16bfb3b93516e144712963fecbc8
Standard model.safetensors SHA-256: 6bb349085512c45fe5cf732209a82a5c5196d2d7a12f0aea16bdb042546dca92
Terminal analysis SHA-256: eb8823e72dbc08a15e78c97449cf6dd7024ab9e7ed0b0f8fd29855ac340e129b
Machine D3 report SHA-256: 92da0b3bf5bd984c579ded700c1b2f9bfe928fe010a5352f65d1a15aea3d48c6
Semantic report SHA-256: 35cc1a87fad2c4f258cfdbd5859d0a0106c0f2c1e8bdd0d6e5ada303a0ffc1e9
Alpha/Transformers parity: 2/2 top-1, maximum logit difference 7.153e-06.

Optimizer-bearing recovery checkpoints and exact restart notes are archived separately in ajaxdavis/alpha-60m-training-checkpoints.

Downloads last month: 8

Safetensors

Model size

57.7M params

Tensor type

F32

Model tree for ajaxdavis/alpha-60m-chat

Base model

ajaxdavis/alpha-60m-base

Finetuned

(1)

this model

ajaxdavis
/

alpha-60m-chat