r4K_anchor_arcmix_qa1/README.md

r4K_anchor_arcmix_qa1 — round 4: anchor arm (ARC-MIX, no upweighting)

Why it exists: the baseline for round 3. Same pool size and source file as r3B_arcmix_qa2x/ (ARC-MIX sampled uniformly, 2.6B-token pool) but with Q&A documents taken once, at their natural share (3.7%). Comparing r3B against this arm isolates the effect of the 2× Q&A upweight; comparing r3A_arcmix_edu_clean/ against it isolates the effect of the educational data.

Results

checkpoint	step (read from file)	ARC-Easy	BLiMP	WikiText-2 byte-ppl	eff (board formula, 62.9M)
ckpt_340k.pt	340000	47.10	76.10	2.3736	75.61
ckpt_360k.pt	360000	47.60	75.60	2.3736	75.61
ckpt_380k.pt	380000	46.89	76.08	2.3696	75.54
ckpt_400k.pt	400000	47.73	75.25	2.3732	75.53

Verdict: Q&A ×2 minus anchor on eff: −0.38 / +0.10 / +0.05 / +0.07; educational data minus anchor: +0.08 / −0.03 / +0.03 / +0.18 (340k-400k). Both are within the seed noise of about 0.3 eff (r4B_arcmix_qa2x_seed1338/). The only repeatable effect is that the educational data makes WikiText-2 slightly worse (by 0.003-0.010 byte-ppl) at all four checkpoints. The anchor itself ends 0.28 eff below the flagship; no arm of rounds 3-4 beats it.

Common setup

• Base model: GoLLeM-v5 64M flagship (v1_muon/, 62.9M parameters, 14 layers, d_model 576, 9 heads, RoPE, SwiGLU, RMSNorm, QK-norm, value residual, Muon optimizer). Every arm of the study starts from the flagship checkpoint at step 320,000 and continues to step 400,000 (80k steps, about 2.6B tokens) with the flagship recipe unchanged (same learning-rate schedule, batch, optimizer state and seed). Only the training data differs.
• Method: two arms trained in parallel from the same checkpoint, evaluated at matching steps (360k / 400k for the first round, 340k to 400k for rounds 3-4) with the same harness; the difference between arms is attributed to the data.
• Evaluation: glint_parity_eval.py in the repository root (fixed version: BLiMP on exactly 67,000 pairs), ARC-Easy test (bare prompt), BLiMP, WikiText-2 test byte-perplexity, eff by the Glint board formula. The step in the tables is read from the checkpoint file, not from its name.
• Reference: flagship v1_muon/ckpt_400k.pt scores ARC-Easy 47.94 / BLiMP 75.83 / byte-ppl 2.3718 / eff 75.81. For scale: a second run that differs only in the random seed (r4B_arcmix_qa2x_seed1338/ vs r3B_arcmix_qa2x/) moved eff by +0.26 / −0.34 / −0.32 / −0.28 at 340k-400k, so run-to-run noise is about 0.3 eff (one seed pair: an order of magnitude, not a precise estimate).
• Status: research checkpoint, not a leaderboard submission. No arm of this study beats the flagship beyond run-to-run noise.
• Format: PyTorch checkpoint dict with model, opt, step, config; train_gpt_ref.py in the repository root rebuilds the model from config.
• Data: ARC-MIX, see the root card of this repository.