SMAXI Community 135M
Experimental small model for extracting task, owner, due, and status fields from short fictional text. Independently fine-tuned from SmolLM2-135M-Instruct. It is not the proprietary SMAXI flagship and contains no flagship weights, private prompts, project data, or application source.
Measured evaluation
48 frozen fictional examples; task/owner values differ from training, but the template is shared. Greedy decoding, maximum 128 output tokens. This narrow test is not a general intelligence, coding, or safety benchmark. The upstream base is a control, not proprietary SMAXI.
| Measure | Upstream base | Community |
|---|---|---|
| Exact JSON object match | 4.2% | 93.8% |
| Valid JSON object | 66.7% | 100.0% |
| Correct fields | 4.2% | 98.4% |
| Median seconds / example | 0.661 | 0.336 |
Device: mps. Timing includes generation and device synchronization, excludes model loading, and is hardware-specific. There is no comparable flagship capability measurement. SMAXI's separate 12/12 software regression result is not a model score.
Training and provenance
- Base revision:
12fd25f77366fa6b3b4b768ec3050bf629380bac - Parameters: 134,515,008
- Full supervised fine-tune: 256 fictional examples, 80 optimizer steps, batch 4, learning rate 0.00002, seed 42.
- Weight SHA-256:
9082ba37e3f2f088eadea5987582339c5e6c27e72af84a1becf583a635eed934 - Dataset SHA-256:
b429fe16da5ee6eb5db697986267d3d488f1b44f3fb6ba073ee820ccd18843e7 - Test SHA-256:
2f2aadc8c37923b9b03c540db122f6c9cff46f3d33c4699d0316529fb11392fe - Actual parameter change verified. No private data or flagship weights used.
This model has no browsing, tools, memory, autonomous execution, privileged APIs, or connection to proprietary SMAXI. It retains capabilities from the public base and can produce incorrect outputs. Open weights can be modified or retrained; limits in the reference runner are not permanent restrictions. Smaller size does not guarantee lower performance than the flagship on every task.
Use
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
name = "Wrooms/SMAXI-Community-135M"
tokenizer = AutoTokenizer.from_pretrained(name, trust_remote_code=False)
model = AutoModelForCausalLM.from_pretrained(name, trust_remote_code=False).eval()
messages = [
{"role": "system", "content": "You are SMAXI Community, a small text-only assistant. Extract task fields into JSON. You have no tools, browsing, private project access, or flagship connection."},
{"role": "user", "content": "Extract task, owner, due, and status as JSON only. Task: label a folder; Owner: Iris; Due: 2027-02-01; Status: todo."}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt", add_special_tokens=False)
if inputs["input_ids"].shape[1] > 1024:
raise ValueError("Reference runner input limit is 1024 tokens")
with torch.inference_mode():
output = model.generate(**inputs, max_new_tokens=128, do_sample=False, pad_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Apache-2.0; see LICENSE and NOTICE for upstream attribution and modifications. This repository distributes downloadable weights. Hosted inference availability is separate.