Skip to content

Instantly share code, notes, and snippets.

@hammer
hammer / xllm-vs-levanter-for-frontier-llm-pretraining.html
Last active September 29, 2026 08:01
IFM vs. Marin: pretraining, data and post-training stacks
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>IFM vs. Marin: pretraining, data and post-training stacks</title>
<style>
:root {
--ink: #1a1a2e; --ink-secondary: #555770; --ink-faint: #8a8a9a;
--accent: #8b2500; --accent-light: #c4530a;
@hammer
hammer / expected-mfu-for-the-535b-a23b-hero-on-megatron-instead-of-l.html
Last active September 23, 2026 13:24
Expected MFU for the 535B-A23B hero on Megatron instead of Levanter
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Expected MFU for the 535B-A23B hero on a Megatron stack instead of Levanter</title>
<style>
:root {
--ink: #1a1a2e; --ink-secondary: #555770; --ink-faint: #8a8a9a;
--accent: #8b2500; --accent-light: #c4530a;
@hammer
hammer / covering-the-competency-map-and-the-task-curriculum-where-ma.html
Created September 22, 2026 06:03
Covering the competency map and the task curriculum: where Marin's RL environments come from
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Covering the competency map and the task curriculum: where Marin&#x27;s RL environments come from</title>
<style>
:root {
--ink: #1a1a2e; --ink-secondary: #555770; --ink-faint: #8a8a9a;
--accent: #8b2500; --accent-light: #c4530a;
@hammer
hammer / looped-transformer-for-marin-s-grug-templates-proposed-issue.html
Created September 20, 2026 18:55
Looped transformer for Marin's grug templates: proposed issue and replication plan for arXiv 2609.19107
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Looped transformer for Marin&#x27;s grug templates: proposed issue and replication plan for arXiv 2609.19107</title>
<style>
:root {
--ink: #1a1a2e; --ink-secondary: #555770; --ink-faint: #8a8a9a;
--accent: #8b2500; --accent-light: #c4530a;
@hammer
hammer / platform-input-support-config.yaml
Last active July 8, 2019 20:28
Configuration organized by ES index; note that this will not parse correctly! It's for purely pedagogical purposes.
# [invalid-]evidence-data
evidences:
gs_output_dir: evidence-files
downloads:
- bucket: otar000-evidence_input/CRISPR/json
output_filename: crispr-{suffix}.json.gz
resource: input-file
subset_key:
- target
- id
@hammer
hammer / mrtarget.data.19.06.yml
Last active July 7, 2019 19:46
Configuration organized by ES index
# relation-data
ddr:
evidence-count: 3
score-threshold: 0.1
# association-data
scoring_weights:
crisp: 1
europepmc: 0.2
expression_atlas: 0.2
@hammer
hammer / Dockerfile
Last active February 24, 2018 20:39 — forked from flying-sheep/Dockerfile
scanpy-locale-setup
FROM ubuntu:17.10
ENV LANG C.UTF-8
RUN apt-get update && \
apt-get install -y python3 python3-pip libxml2-dev zlib1g-dev wget git cmake && \
apt-get clean
RUN apt-get install -y python3-numpy # not necessary in scanpy 0.2.9.2/0.2.10
@hammer
hammer / cosmic_sigs.R
Last active January 9, 2017 21:19
Exploring COSMIC signatures
# deconstructSigs signatures.cosmic
library("deconstructSigs")
library("tibble")
signatures.cosmic.tidy <- tibble::rownames_to_column(as.data.frame(t(signatures.cosmic)))
# COSMIC http://cancer.sanger.ac.uk/cancergenome/assets/signatures_probabilities.txt
cosmic.current <- read.delim("http://cancer.sanger.ac.uk/cancergenome/assets/signatures_probabilities.txt", check.names = F)
cosmic.current.clean <- cosmic.current[ , colnames(cosmic.current) != ""]
cosmic.current.tidy <- cosmic.current.clean[,names(cosmic.current.clean) not in c("Substitution Type","Trinucleotide")]
@hammer
hammer / seqs2bed.py
Created January 11, 2016 22:20
Convert DeepSEA training sequences a BED file
import h5py
# HDF5 file with two arrays: 'trainxdata' (samples) and 'traindata' (labels)
INFILE_SAMPLES = ''
INFILE_REFERENCE_FASTA = ''
OUTFILE_FASTA = 'deepsea_train10k.fa'
OUTFILE_BED = 'deepsea_train10k.bed'
def onehot2base(onehot):
if onehot == [1,0,0,0]:
type sample = {
x : float;
y : float
}
type model = {
theta : float;
beta : float
}