quantal-classroom-1.6

The ring-of-teachers quantal brain. A 1.7B thresholded-ternary student distilled from a faculty of two tokenizer-identical teachers (Qwen3-8B + Qwen3-14B), trained with the deployed-forward BitNet b1.58 quantizer that the ayeOS Rust runner executes unchanged.

Why this number matters

run val (masked CE) note
CE-only baseline 2.1469 the line to beat
single-teacher KL (old stack, vast H100) 2.1369 old harness
harness gate 2.1369 → 2.1369 same weights, new stack: the gap is training
v2 single-teacher (HF Jobs H200) 1.8166 new stack, real gain
classroom-1.6 (8B+14B ring) 1.6120 the ring beats every single teacher

The harness gate is the honesty check: the archived 2.1369 checkpoint re-evaluated on the new stack returns 2.1369 exactly. Same weights, same number — so every improvement below 2.1369 is training, not machinery.

The quantizer (unchanged from training to Rust)

  • thresholded-ternary: per-G=64-group scale = mean(|w|), band 0.5·scale, codes in {-1, 0, +1}
  • deployed-forward: the Rust runner reads codes+scales and computes (code − 1) · scale — the training forward is bit-identical to inference
  • fixed-point level rule (accepted r8/r10): a_{n+1} = mean(|w| ≥ 0.5·a_n) converges; C(a) is monotone and bounded, so k ≥ 1 in every group — no group ever collapses to all-zeros (proof by monotone map, not just the pass-1 interval bound)

The classroom (ring-of-teachers)

  • faculty: Qwen3-8B + Qwen3-14B, both tokenizer byte-identical to the student (vocab 151,643)
  • loss: 0.5·CE + 0.5·mean_f KL(student ∥ teacher_f), β ramped from 0 over 2 epochs
  • consensus: the student binds to the faculty mean teaching signal, not any single teacher's errors

Files

  • quantal_model.safetensors — full-precision BitLinear weights (the deployed-forward quantizer is applied at export/inference time)
  • config.json, tokenizer.* — Qwen3-1.7B architecture
  • the ayeOS ternary capsule export (m*.json + index.json) ships on request / via the pocoo demo

The harder question (provenance)

This checkpoint exists because the box that held its predecessor died, and the story of which weights are which is recorded honestly:

  • 2d54a10f — old blob, relabel pending
  • 21294c68 — the 2.1469 CE-only export (published, the other arm)
  • 834dc609 — pocoo demo blob
  • ckpts-classroom/quantal-long-best.safetensors — this run's best

Integrity preserves what was asserted; it does not establish that the assertion was true. The KL arm (the winner) is now published here so the provenance gap cannot recur.

Downloads last month
-
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for PeetPedro/quantal-classroom-1.6

Finetuned
Qwen/Qwen3-1.7B
Finetuned
(974)
this model

Dataset used to train PeetPedro/quantal-classroom-1.6

Space using PeetPedro/quantal-classroom-1.6 1

Article mentioning PeetPedro/quantal-classroom-1.6