Helw150

Four intuitive high-quality Harrier mixtures

These are four new 18.75T-token, two-phase priors for d768 seed 0. Phase 0 is 15T tokens (80%); phase 1 starts at step 8448 and is 3.75T tokens (20%). The quality assumption is monotonic: q4 is best, followed by q3, q2, q1, and q0.

The recipes are intentionally prior-driven. Earlier experiments inform the cumulative epoch cap: every cXXqY source stays at or below four epochs across both phases. Use the CSV weights verbatim; its 400 weight columns are the complete phase-0 and phase-1 mixtures.

Recipes

Family shares are ordered code_agent / science_math / knowledge / low_value / general. Quality shares are ordered q0 / q1 / q2 / q3 / q4.

Helw150

Variance-normalized Harrier mixture objective

Summary

We score mixture experiments in units of training-seed variation measured from ten Marin proportional runs. Lower scores are better.

For each target task, a one-standard-deviation BPB improvement contributes -1. For each guardrail task, regressions up to one proportional-seed standard deviation are free. A two-standard-deviation guardrail regression contributes +1.

This is the variance-normalized form of the original hinge objective. Target tasks are stable BPB proxies for capabilities in the Artificial Analysis Intelligence Index v4.1.1. The current results and per-run mixture metadata are in the harrier_18t75_d768 dataset subset.

Helw150

Grug step-630 · MATH500 full-context results

Metric	Value
Accuracy	54.2%
Solved	271 / 500
Example limit	none
Model context	32,768 tokens
Generation allowance	28,672 tokens

Helw150

SimpleRL-Zoo PPL viz

Base-model PPL vs post-RL math accuracy (issue #6096).

Helw150

"""Build a single-file static HTML mixture explorer from the swarm CSVs.

Output: docs/index.html (data inlined, Plotly via CDN).
Open it directly with `file://` or serve with `python -m http.server -d docs/`.
The docs/ folder is what GitHub Pages serves.

Per-task fits use Calvin's DSP canonical form:
  e_p,d = c_p,d * w_p,d
  z_d   = e_0,d + gamma * e_1,d
  y_hat(w) = -b0 + sum_d a_d (1 - exp(-rho_d z_d))

Helw150

GCS Storage Report

Generated: 2026-05-26T20:00:58Z

Overview

Metric	Value
Total Objects	321.4M
Total Size	3,049.70 TB

Helw150

Helw150

GCS Storage Report

Generated: 2026-05-23T00:30:00Z

Overview

Metric	Value
Total Objects	338.6M
Total Size	3,110.95 TB

Helw150

GCS Storage Report

Generated: 2026-05-19T05:03:56Z

Overview

Metric	Value
Total Objects	233.3M
Total Size	3,010.88 TB

Helw150

GCS Storage Report

Generated: 2026-05-19T03:51:40Z

Overview

Metric	Value
Total Objects	231.0M
Total Size	3,009.25 TB