Record: PR1855/PR1953 base + Progressive context growth (val_bpb: 1.05759, 3-seed)

[simonbissonnette]

Record candidate: SP8192 CaseOps + Progressive 3k Context Growth + Short-Doc Score-First TTT

val_bpb: 1.05759 (3-seed mean, std 0.00034) | val_loss: 2.31441 nats (std 0.00075) | 15.98 MB max | 8xH100 SXM | 600s train / 600s eval

Improvement over merged PR #1855 leaderboard record (1.06107587 BPB):
-0.00348 BPB / -0.00762 nats

This stacks a progressive training-context schedule and a short-document TTT schedule on top of the late-April CaseOps/SP8192/LQER/SparseAttnGate/BOS-fixed SmearGate lineage. The direct leaderboard comparison is PR #1855, which is the current merged leader used here as the baseline.

Results

Seed	Steps	ms/step	Train ms	Pre-quant BPB	Quant BPB	Post-TTT BPB	TTT eval s	Artifact bytes
42	4,888	121.9	596,025	1.05993108	1.06833072	1.05740567	572.4	15,981,945
314	4,882	122.1	595,976	1.05975470	1.06832443	1.05730104	489.9	15,984,387
0	4,884	122.0	596,022	1.06072266	1.06902034	1.05807084	493.5	15,981,122
Mean	4,884.7	122.0	596,008	1.06013615	1.06855850	1.05759252	518.6	15,982,485

3-seed population std: 0.00034091 BPB / 0.00074604 nats.

All included seeds are under the 16,000,000-byte artifact cap and the 600s train/eval budgets as logged. The maximum artifact is 15,984,387 bytes and the maximum validation-data TTT pass is 572.4s.

Full validation coverage

All three logs evaluate the full CaseOps validation shard target set:

Seed	val_tokens	target_tokens
42	47,853,343	47,853,343
314	47,853,343	47,853,343
0	47,853,343	47,853,343

The training script explicitly keeps the validation tail via EVAL_INCLUDE_TAIL=1. This avoids the older multiple-of-context truncation and makes the standard diagnostic eval and quantized TTT eval agree on the same target count.

The tokenizer, CaseOps transform, training shards, validation shard, and byte sidecar format are the same canonical HF-hosted CaseOps export used by the merged PR #1855 setup. If a reviewer already has the clean #1855/HF CaseOps data staged, those same staged shards can be reused here. The included tokenizer/prep files are present only to make this submission self-contained; the preferred reproduction path is to download the canonical HF CaseOps export directly.

What changed vs PR #1855

This submission keeps the same overall 11-layer SP8192 CaseOps recurrent-transformer family as PR #1855, then adds the following levers:

Lever	Setting	Purpose
Progressive train context	TRAIN_SEQ_SCHEDULE=1024@0.100,2048@0.700,3072@1.000	Train cheaply at 1k early, move to 2k for most of training, then finish at 3k context.
Final/eval context	TRAIN_SEQ_LEN=3072, EVAL_SEQ_LEN=3072, TTT_EVAL_SEQ_LEN=3072, EVAL_STRIDE=1536	Extend the final model and TTT scoring context beyond 2k without the 4k eval-time cost.
Long-context TTT mask	TTT_MASK=no_qv, TTT_Q_LORA=0, TTT_V_LORA=0	Keep K/O/MLP LoRA adaptation while removing Q/V adapters that were less helpful at longer context.
TTT local LR	TTT_LOCAL_LR_MULT=0.75	Slightly softer per-document LoRA adaptation.
Short-doc score-first chunks	TTT_SHORT_SCORE_FIRST_STEPS=256:8,2000:24, default chunk 48	Use smaller score-before-update chunks for short documents, preserving causality while improving adaptation.
TTT phases	PHASED_TTT_NUM_PHASES=1, PHASED_TTT_PREFIX_DOCS=2500	Single score-first phased pass with a 2500-doc prefix budget.
QK gain	QK_GAIN_INIT=5.25	Public long-context sweep result from the PR #1953 lineage.
Compression/quant stack	COMPRESSOR=pergroup, AWQ-lite, asymmetric logit rescale	Inherited from public late-April quantization/compression work stacked on the PR #1855 base.

The short-doc TTT schedule does not train on future validation tokens. It only changes the chunk granularity used inside the existing score-before-update loop: each chunk is scored first, then the LoRA update is applied for future chunks.

Architecture and training stack

Component	Setting
Model	11 layers, 512d, 8 query heads, 4 KV heads, MLP 4x
Tokenizer/data	SP8192 CaseOps lossless caps with byte sidecar accounting
RoPE	Partial RoPE, 16 dims
Recurrence	Layers 3-5 looped, enabled at frac=0.35
Parallel decoder	Parallel lane from layer 8, mean final lane
XSA	All 11 layers
Gates	BOS-fixed SmearGate, SparseAttnGate with gate_window=12, scale 0.5
Optimizer	Muon on matrix params, Adam on embedding/scalars, BETA2=0.99
EMA	ema_decay=0.9965
Quantization	GPTQ int6 matrices, int7 embeddings, LQER asymmetric rank-4 correction
GPTQ reserve	GPTQ_RESERVE_SECONDS=4.0; logs show gptq:reserving 4s, effective=596000ms
Compression	Per-group compression
TTT	Quantized phased LoRA TTT, score-first, no_qv mask, short-doc chunk schedule

Compliance notes

• Artifact cap: all seeds <= 15,984,387 bytes.
• Training wallclock: all training loops stop around 596.0s with GPTQ_RESERVE_SECONDS=4.0; GPTQ hessian collection is logged immediately after (67 Hessians in 4.1s) for transparency.
• Eval wallclock: all validation-data TTT passes are <= 572.4s. The ttt_lora:compile warmup uses random tokens and no validation data; it is logged separately from total_eval_time.
• Score-before-update: quantized_ttt_phased scores each chunk before applying that chunk's LoRA update. The short-doc schedule only changes chunk size.
• Full validation targets: val_tokens == target_tokens == 47853343 in all included logs.
• No validation data in training: training uses only training shards. TTT accesses validation documents left-to-right under the score-first rule.
• No external cache or direct memorization: no SLOT, n-gram cache, PPM mixture, logit bias table, or validation-derived precomputation.
• Original-byte BPB: CaseOps byte sidecar accounting is preserved.

Reproduction

Install the dependencies in requirements.txt. FlashAttention 3 and the lrzip system binary are noted there because they require separate install paths.

This submission uses the clean canonical CaseOps SP8192 export hosted on Hugging Face. The logs were produced from a 50,000-document validation split with 80 training shards (train_shards: 80, ttt_phased: total_docs:50000, and val_tokens == target_tokens == 47853343 in every included log).

Preferred data setup:

python3 - <<'PY'
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="romeerp/parameter-golf-caseops-v1",
    repo_type="dataset",
    local_dir="./data/datasets/fineweb10B_sp8192_caseops",
    allow_patterns=[
        "datasets/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved/*",
        "datasets/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model",
    ],
    max_workers=8,
)
PY

Then set:

DATA_PATH=./data/datasets/fineweb10B_sp8192_caseops/datasets/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved
TOKENIZER_PATH=./data/datasets/fineweb10B_sp8192_caseops/datasets/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model

Fallback local rebuild: if the HF export is unavailable, rebuild from the canonical docs_selected.jsonl with the included prepare_caseops_data.py, lossless_caps.py, and tokenizer. Use --val-docs 50000 and write into a fresh output directory. The prep script now defaults to 50,000 validation docs and refuses to write over existing fineweb_*.bin shards unless --overwrite is passed, to avoid accidentally mixing stale validation shards with a new train split.

Run one seed at a time, replacing DATA_PATH and TOKENIZER_PATH with the staged CaseOps paths:

for SEED in 42 314 0; do
  NCCL_NET=Socket \
  DATA_DIR=./data \
  DATA_PATH=./data/datasets/fineweb10B_sp8192_lossless_caps_caseops_v1_reserved \
  TOKENIZER_PATH=./data/tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model \
  CASEOPS_ENABLED=1 \
  VOCAB_SIZE=8192 \
  ITERATIONS=20000 \
  MAX_WALLCLOCK_SECONDS=600 \
  EVAL_INCLUDE_TAIL=1 \
  TRAIN_SEQ_LEN=3072 \
  ROPE_TRAIN_SEQ_LEN=3072 \
  TRAIN_SEQ_SCHEDULE=1024@0.100,2048@0.700,3072@1.000 \
  TRAIN_SEQ_SCHEDULE_MODE=wallclock \
  SEQ_CHANGE_WARMUP_STEPS=32 \
  EVAL_SEQ_LEN=3072 \
  EVAL_STRIDE=1536 \
  TTT_ENABLED=1 \
  TTT_EVAL_SEQ_LEN=3072 \
  TTT_BATCH_SIZE=24 \
  TTT_CHUNK_SIZE=48 \
  TTT_SHORT_SCORE_FIRST_ENABLED=1 \
  TTT_SHORT_DOC_LEN=2000 \
  TTT_SHORT_CHUNK_SIZE=24 \
  TTT_SHORT_SCORE_FIRST_STEPS=256:8,2000:24 \
  TTT_LORA_RANK=80 \
  TTT_LORA_LR=0.0001 \
  TTT_LOCAL_LR_MULT=0.75 \
  TTT_MASK=no_qv \
  TTT_Q_LORA=0 \
  TTT_V_LORA=0 \
  TTT_WEIGHT_DECAY=0.5 \
  TTT_BETA2=0.99 \
  PHASED_TTT_PREFIX_DOCS=2500 \
  PHASED_TTT_NUM_PHASES=1 \
  WARMDOWN_FRAC=0.85 \
  BETA2=0.99 \
  QK_GAIN_INIT=5.25 \
  SPARSE_ATTN_GATE_ENABLED=1 \
  SPARSE_ATTN_GATE_SCALE=0.5 \
  GATED_ATTN_QUANT_GATE=1 \
  SMEAR_GATE_ENABLED=1 \
  GATE_WINDOW=12 \
  FUSED_CE_ENABLED=1 \
  MATRIX_LR=0.026 \
  MIN_LR=0.1 \
  GRAD_CLIP_NORM=0.3 \
  EMBED_BITS=7 \
  EMBED_CLIP_SIGMAS=14.0 \
  MATRIX_CLIP_SIGMAS=12.85 \
  ATTN_CLIP_SIGMAS=13.0 \
  MLP_CLIP_SIGMAS=11.5 \
  LQER_ENABLED=1 \
  LQER_RANK=4 \
  LQER_TOP_K=3 \
  LQER_FACTOR_BITS=4 \
  LQER_ASYM_ENABLED=1 \
  LQER_ASYM_GROUP=64 \
  AWQ_LITE_ENABLED=1 \
  AWQ_LITE_BITS=8 \
  AWQ_LITE_GROUP_TOP_K=1 \
  AWQ_LITE_GROUP_SIZE=64 \
  ASYM_LOGIT_RESCALE=1 \
  GPTQ_RESERVE_SECONDS=4.0 \
  GPTQ_CALIBRATION_BATCHES=16 \
  COMPRESSOR=pergroup \
  VAL_LOSS_EVERY=0 \
  SEED=$SEED \
  torchrun --standalone --nproc_per_node=8 train_gpt.py \
      > train_seed${SEED}.log 2>&1
done

Included files

• train_gpt.py - full training/eval script used for the logs.
• train_seed42.log, train_seed314.log, train_seed0.log - full per-seed logs.
• submission.json - structured metadata and per-seed results.
• README.md - this file.
• requirements.txt - Python dependencies plus notes for FA3 and lrzip.
• prepare_caseops_data.py - fallback CaseOps dataset/token/byte-sidecar preparation; defaults to the canonical 50,000-doc validation split and refuses mixed/stale output directories by default.
• lossless_caps.py - reversible CaseOps transform, same as the PR Record: SP8192 + LQER + Sparse Attn Gate + BOS-Fixed SmearGate + 9-Hparam Greedy Stack — val_bpb 1.06108 (3-seed mean) #1855 CaseOps setup.
• tokenizers/fineweb_8192_bpe_lossless_caps_caseops_v1_reserved.model - SentencePiece tokenizer used by the logs; identical CaseOps tokenizer lineage as PR Record: SP8192 + LQER + Sparse Attn Gate + BOS-Fixed SmearGate + 9-Hparam Greedy Stack — val_bpb 1.06108 (3-seed mean) #1855.

Lineage and credits

This submission is a stack on top of the public CaseOps/SP8192 record lineage:

• PR Record: SP8192 + LQER + Sparse Attn Gate + BOS-Fixed SmearGate + 9-Hparam Greedy Stack — val_bpb 1.06108 (3-seed mean) #1855 by @codemath3000 - merged leaderboard record and direct comparison baseline.
• PR Record: V22 = V21 + PR #1953 levers + EVAL_SEQ_LEN=2816 -- val_bpb 1.05877 (3-seed mean, all strict <600s) #1945 / PR Record: PR #1855 base + activation-aware GPTQ mixed precision - val_bpb 1.06081 (3-seed mean) #1908 / PR Record: SP8192 #1855 Base + Asymmetric Logit Rescale + AWQ-lite — val_bpb 1.05971 (3-seed mean, full val) #1923 public late-April quantization stack - AWQ-lite and asymmetric logit rescale lineage.
• PR Record: PR #1945 base + 2560 long-context + no_qv TTT mask + TTT LR 0.75 + QK_GAIN 5.25 — val_bpb 1.05855 (3-seed mean) #1953 - long-context/no_qv/QK-gain sweep ideas.
• PR Record: PR #1787 base + Smear Gate + LQER Asym — val_bpb 1.06157 #1797 by @dexhunter - SmearGate and LQER asymmetric rank-4 lineage.
• PR Record: PR #1736 + Polar Express NS + MIN_LR + Sparse Attn Gate + Fused CE + PR #1767 TTT — val_bpb 1.06335 #1787 by @nprime06 - Polar Express Muon, MIN_LR, SparseAttnGate, fused CE.
• PR Record: SP8192 + CaseOps + GatedAttn + QuantGate + Loop45 + PhasedTTT — val_bpb 1.06549 #1736 and PR Record: CaseOps Tokenizer + Tapered WD - val_bpb 1.0678 (3-seed mean) #1729 by @dexhunter / @romeerp - CaseOps integration and byte sidecar accounting.
• PR RECORD: SmearGate + Attention Output Gate + Legal TTT | val_bpb=1.07139 #1667 by @MarioPaerle - SmearGate lineage.
• PR Record: VarLen Attention + Fused MLP + Multi-Phase Global SGD TTT — val_bpb 1.07193 (3-seed mean) #1626 / PR Record: VarLenAttn + PhasingTTT - val_bpb 1.0728 (3-seed mean) #1610 - phased score-first TTT lineage.
• Issue A Field Guide to Valid Submissions #1017 by @cocohearts - score-first validation criteria.

The new contribution here is the combination of progressive 3k train/eval context growth with the short-document score-first TTT chunk schedule, while preserving the full validation target count and staying under the artifact/eval budgets.