soasme

  ┌────────────┬──────────────────────┬──────┬──────┬────────┬───────────────┬────────────────────┐
  │    spec    │   distinct states    │ TLC  │ tpp  │ tla-rs │ tpp vs tla-rs │   memory tpp /     │
  │            │                      │      │      │        │               │       tla-rs       │
  ├────────────┼──────────────────────┼──────┼──────┼────────┼───────────────┼────────────────────┤
  │ closure    │                65536 │ 2.25 │ 0.47 │   4.02 │          8.6× │         21 / 48 MB │
  ├────────────┼──────────────────────┼──────┼──────┼────────┼───────────────┼────────────────────┤
  │ counters   │               823543 │ 3.79 │ 0.76 │  16.83 │           22× │       186 / 654 MB │
  ├────────────┼──────────────────────┼──────┼──────┼────────┼───────────────┼────────────────────┤
  │ liveness   │                38880 │ 4.59 │ 0.23 │   0.79 │          3.4× │        104 / 77 MB │

soasme

soasme

{"id":"0ae06a70eb82464ab2833e5d38f3c0a7","timestamp":1788852607.755248,"type":"session_info","created_at":1788852607.755248,"cwd":"/Users/soasme/github.com/soasme/rio","skill":"rio-coding"}
{"id":"4ab27d02ad4745a99d453cc6a3ad06f1","parent_id":"0ae06a70eb82464ab2833e5d38f3c0a7","timestamp":1788852607.75527,"type":"leaf","entry_id":"0ae06a70eb82464ab2833e5d38f3c0a7"}
{"id":"be8f78350190464a84f31236a9d7b10a","parent_id":"0ae06a70eb82464ab2833e5d38f3c0a7","timestamp":1788852617.606848,"type":"turn","observation":"work on https://github.com/soasme/rio/issues/45 and create a PR.","state":{"goal":"","plan":[],"findings":{},"files":{},"cwd":"/Users/soasme/github.com/soasme/rio","environment":{"cwd":"/Users/soasme/github.com/soasme/rio","os":"Darwin","shell":"/bin/zsh","date":"2026-09-08"},"blockers":[],"last_error":null,"scratch":{}}}
{"id":"c1792765fca141aca0984d5dfaadb43d","parent_id":"be8f78350190464a84f31236a9d7b10a","timestamp":1788852617.6075158,"type":"leaf","entry_id":"be8f78350190464a84f31236a9d7b10a"}
{"id"

soasme

{"id":"cb62dadb80734b54a67bcdc60909e834","timestamp":1788818943.5539289,"type":"session_info","created_at":1788818943.5539289,"cwd":"/Users/soasme/github.com/soasme/rio","skill":"rio-coding"}
{"id":"1e8f1ed82fce497faf84095d58f7f122","parent_id":"cb62dadb80734b54a67bcdc60909e834","timestamp":1788818943.5539541,"type":"leaf","entry_id":"cb62dadb80734b54a67bcdc60909e834"}
{"id":"c70a3f701faa43e98e43940159685be1","parent_id":"cb62dadb80734b54a67bcdc60909e834","timestamp":1788819885.45872,"type":"model_change","model":"claude-sonnet-4-6","provider":"anthropic"}
{"id":"1cf09953ed044998990df76f3175724a","parent_id":"c70a3f701faa43e98e43940159685be1","timestamp":1788819885.458742,"type":"thinking_level_change","thinking_level":"medium"}
{"id":"58679692031040b8a7660458f6d1aeee","parent_id":"1cf09953ed044998990df76f3175724a","timestamp":1788819885.458776,"type":"leaf","entry_id":"1cf09953ed044998990df76f3175724a"}
{"id":"0f42579132ab47a5aea328a5f71c91fc","parent_id":"1cf09953ed044998990df76f3175724a","timestamp":178881

soasme

"""Demo: run RMSNorm on a random batch and check it against torch.nn.RMSNorm."""

import torch
import torch.nn as nn

from rms_norm import RMSNorm


def demo() -> None:
    torch.manual_seed(0)

soasme

train_tokenizer.py

A byte-level BPE tokenizer trainer/evaluator.

Changes from the original:

• CLI-parametrized (no hardcoded paths, vocab size, or special-token count).
• Split into small, independently testable functions instead of one script.
• train and eval are separate subcommands.
• Ships with a 100-row multilingual sample corpus (English, Spanish, French,

soasme

"""
micromlp: a single file, no-dependency neural net that predicts California
housing prices — the dataset from chapter 2 of Hands-On Machine Learning.

Downloads the housing CSV, caches it locally, engineers a small feature
set, and trains a 2-layer MLP with a from-scratch autodiff engine and
plain gradient descent to predict median_house_value.

Usage: python3 micromlp.py
"""

soasme

#!/usr/bin/env python3
"""
Demo of a GitHub credential-broker proxy: injects a short-lived `gh auth
token` into outgoing requests, so a sandboxed VM never holds the real
credential. This is a demo only — it does not implement any policy
checks (allow/deny by repo, branch, or push type). Add that at the
proxy layer before using this for anything real.
"""
import base64
import os

soasme

---
name: small-model-finetuning
description: Build, evaluate, fine-tune, quantize, and deploy small language models for narrow production tasks using an eval-first dataset factory workflow.
version: 1.0.0
language: en
tags:
  - small-language-models
  - fine-tuning
  - qlora
  - lora

soasme

The New Engineering Teams

How AI coding agents change the way teams should work

---

Opening

Most engineering orgs were designed for one constraint: