Add curated checkpoints, checkpoints/README.md and mixed French corpus shards

checkpoints/README.md

@@ -0,0 +1,123 @@
+# Checkpoints
+
+This is a curated set, not every checkpoint that was trained. **None of these models writes coherent French.** Use them to
+reproduce the measurements in the top-level README, to inspect the model, or to continue training.
+
+All `.pt` files are ordinary ATOM checkpoints (`AtomNativeModel` format, d=32, 479,548 parameters). They load with
+`tools/chat_atom_native.py`, `tools/run_atom_native.py --resume` and `tools/train_batched.py --init`. The checkpoints carry
+optimizer state, which is why they are ~3.8 MB. The GPU checkpoint is ~10 MB because it also carries the per-stream runtime state for 512 streams.
+
+Metric: next-byte cross-entropy in nats/byte (normal / field-alone), measured by the run's own progress script. "Dialogue" and "varied" are
+the true held-out sets `data/heldout/val_dialogue.txt` and `data/heldout/val_varied.txt` (4 and 6 windows × 512 bytes). "val2048" is the old
+4-window validation inside the template corpus.
+
+## Recommended
+
+| file | size | run | step | val numbers (source) | use it for |
+|---|---|---|---|---|---|
+| `byte_tick_bptt_k128_mix/atom_native_step_1720000_d32_bptt_k128_mix.pt` | 3,816,181 B | K128 BPTT continued on the 7.4 MB mixed French corpus | 1,720,000 | not evaluated yet (the run was still training at upload time) | **latest**; default for chat and for resuming |
+| `byte_tick_bptt_k128_mix/atom_native_step_1680000_d32_bptt_k128_mix.pt` | 3,816,181 B | same | 1,680,000 | dialogue **1.0053** / 1.0317, varied 2.0600 / 2.0768, val2048 1.0693 / 1.0962 (`docs/artifacts/mix_long_memory/bptt_k128_mix_progress.tsv`) | **best held-out dialogue** of the run |
+| `byte_tick_bptt_k128_mix/atom_native_step_1600000_d32_bptt_k128_mix.pt` | 3,816,181 B | same | 1,600,000 | dialogue 1.0212 / 1.0337, varied **2.0581** / 2.0701, val2048 1.0961 / 1.1227 (same TSV) | **best held-out varied (book) text** |
+| `byte_tick_bptt_fresh/atom_native_step_1232000_d32_bptt_fresh.pt` | 3,815,607 B | fresh init + BPTT K=64 on the template corpus (`data/corpus_fr_medium*.txt`) | 1,232,000 | val2048 **0.3560** / 0.3945 (`docs/artifacts/decfix_bptt/bptt_fresh_progress.tsv`) | best fresh K64 row; recites template lines |
+| `byte_tick_bptt_fresh_k128/atom_native_step_1104000_d32_bptt_fresh_k128.pt` | 3,816,649 B | fresh init + BPTT K=128 on the template corpus | 1,104,000 | val2048 **0.3954** / 0.4653 (`bptt_fresh_k128_progress.tsv`); held-out dialogue 1.2856 / 1.2914, varied 6.7716 / 6.5909 (`docs/artifacts/mix_long_memory/k128_1104k_before_mix.tsv`) | best fresh K128 row; starting point of the mixed-corpus run |
+| `gpu_smoke_b512/atom_native_step_4000_smoke_b512.pt` | 10,096,441 B | GPU smoke test: batched B=512 on an RTX 5090, `--init` from the K128-mix checkpoint 1,616,000, 4,000 ticks (2.05 M bytes) | 4,000 (batched ticks) | val CE 1.2957 → **1.2307** (field-alone 1.3174 → 1.2644) from tick 1000 to 4000 (`gpu_smoke_b512/progress.tsv`; `train_batched.py` val protocol, 4 windows over the held-out files, so not directly comparable to the TSV above) | proof that GPU checkpoints load and chat on CPU |
+
+Each directory also contains the run's `run_metrics_*.json` (or `progress.tsv` for the GPU run) exactly as the trainer wrote it.
+In `run_metrics_*.json`, `shard_paths` show absolute paths from the training machine; they are informational only.
+
+## `RELEASE/` (historical)
+
+These are 42 snapshots from **before** the root-cause fixes. They use the old recipe, with no BPTT through the field, and were trained mostly on the template corpus: a d16 120k tip,
+the d32 s21 series (32k … 496k), and the 2.90M / 2.925M dialogue/hard-readout tips. See `RELEASE/README.md` and the `*.run_metrics.json` next to each file.
+They load and run, but they are not recommended: for example, `atom_native_2_925M_hard25k.pt` emits byte noise. They are kept for comparison with the history in `docs/`.
+
+## Install (CPU)
+
+```bash
+git lfs install
+git clone https://huggingface.co/thefinalboss/atom-ai
+cd atom-ai
+python3 -m venv .venv
+.venv/bin/pip install --index-url https://download.pytorch.org/whl/cpu torch   # CPU wheel; any torch >= 2.1 works
+.venv/bin/pip install -r requirements.txt                                      # numpy (+ pytest)
+```
+
+Do not use `pip install -e .`. The legacy `pyproject.toml` pulls in `transformers`/`datasets`, which ATOM never imports.
+All tools are run from the repo root with `PYTHONPATH=.`. Tested with Python 3.13 and torch 2.14.1+cpu.
+
+## Chat on CPU
+
+```bash
+PYTHONPATH=. .venv/bin/python tools/chat_atom_native.py \
+  --checkpoint checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1720000_d32_bptt_k128_mix.pt \
+  --prompt "Bonjour, comment ça va ?" --device cpu
+```
+
+Useful flags: `--temperature 0.7 --top-k 8` (defaults), `--deterministic` (greedy; currently loops on `Je suis d'accord ?`),
+`--max-length 256`, `--interactive`, `--no-role-prime` (do not wrap as `Utilisateur:/Assistant:`), `--legacy-gen-labels` (reproduces the
+pre-fix decode bug, for comparison only). One reply takes about 2 s on one CPU thread.
+
+## Resume training on CPU (single stream, exact recipe of the K128-mix run)
+
+This is one 8,000-step chunk from the latest checkpoint. The stream offset is `(step − 1,104,000) mod 7,394,592` bytes, so for 1,720,000 it is 616,000.
+
+```bash
+PYTHONPATH=. OMP_NUM_THREADS=1 .venv/bin/python -u tools/run_atom_native.py \
+  --stream --data-glob 'data/corpus_fr_mix_[0-9]*.txt' --loop-shards --chunk-bytes 65536 \
+  --stream-skip-bytes 616000 \
+  --resume checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1720000_d32_bptt_k128_mix.pt \
+  --output-dir checkpoints/my_k128_mix --checkpoint-name atom_native_step_1728000_my_k128_mix.pt --steps 8000 \
+  --d-model 32 --n-modes 32 --n-atoms-max 64 --max-span-bytes 1 \
+  --field-obligatory-hard --no-enable-merge --no-payload-copy --last-atom-readout \
+  --efference-every 0 --learning-rate 2e-3 --surface-learning-rate 2e-3 --seed 20260913 \
+  --atom-flush-every 64 --episode-length 8000 --episode-reset --slow-every 1 \
+  --field-loss-weight 0.05 --field-contrast-weight 0 --allow-parallel-train \
+  --last-atom-dropout 0.5 --field-leak 0.05 --field-bptt 128 --field-leak-learnable
+```
+
+Single-stream CPU speed is about 100 steps (= bytes) per second on one core, and more threads do not help at d=32.
+`scripts/train_bptt.sh` chains such chunks automatically and resumes from the newest checkpoint in `OUT`. Its variables are documented in
+`docs/MIX_DATA_LONG_MEMORY.md`. Note that it (like several older scripts) starts with `cd /workspace/exports/atom-ai`, the original machine's path, so edit that line
+before using it elsewhere. Measure a checkpoint on the held-out sets with
+`docs/artifacts/mix_long_memory/tsv_mix.py <out.tsv> <ckpt.pt> <run_metrics.json>`.
+
+## Batched training (GPU, or CPU for testing): `tools/train_batched.py`
+
+B independent byte streams run in lockstep with the same recipe (K=128 BPTT, learnable leak, LR 2e-3). 1 tick = B bytes.
+The full runbook is in `docs/BATCHED_TRAINING.md`.
+
+GPU box setup (creates `.venv`, installs a CUDA torch, runs the CUDA==CPU tests and a benchmark):
+
+```bash
+bash scripts/setup_gpu_box.sh
+# if torch cannot see the GPU: TORCH_INDEX=https://download.pytorch.org/whl/cu128 bash scripts/setup_gpu_box.sh
+```
+
+Continue the latest CPU checkpoint on the GPU (resumable; re-running the same command resumes from the newest checkpoint in `OUT`):
+
+```bash
+mkdir -p logs/batched
+nohup env DATA_GLOB='data/corpus_fr_mix_[0-9]*.txt' VAL_GLOB='data/heldout/*.txt' B=512 DEVICE=cuda \
+  INIT=checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1720000_d32_bptt_k128_mix.pt \
+  TOTAL=4000000 CKPT_EVERY=8000 \
+  bash scripts/train_batched.sh >> logs/batched/master.log 2>&1 < /dev/null &
+# monitor: tail -f logs/batched/master.log ; column -t -s$'\t' checkpoints/batched_b512/progress.tsv | tail
+# stop:    kill $(cat checkpoints/batched_b512/driver.pid)
+```
+
+The same run can be started directly, e.g. as a small CPU test (checked for this upload with `-B 4 --total-ticks 128`):
+
+```bash
+PYTHONPATH=. .venv/bin/python -u tools/train_batched.py \
+  --data-glob 'data/corpus_fr_mix_[0-9]*.txt' --val-glob 'data/heldout/*.txt' \
+  --init checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1720000_d32_bptt_k128_mix.pt \
+  -B 8 --device cpu --threads 1 --output-dir checkpoints/bt_test --suffix bt_test \
+  --total-ticks 400 --ckpt-every 200 --log-every 50
+```
+
+Measured throughput: RTX 5090 at K=128 ran ~85.9k bytes/s at B=1024, ~125k at B=1536 and ~145k at B=1792 (CUDA peak memory 14.4 / 21.5 / 25.1 GB;
+`docs/artifacts/batched_training/bench_gpu_quickpod_rtx5090*.json`). On one CPU thread, B=32 / B=128 give ×19 / ×41 the bytes per CPU-second of the legacy single-stream loop.
+`--init` starts the batched tick counter at 0 (the loaded step is recorded). Batched checkpoints are written as `atom_native_step_<tick>_<suffix>.pt`,
+and they chat on CPU with the same command as above.
+
+For a multi-GB French corpus see `scripts/fr_big/` (download + build of Wikipedia FR, Gutenberg FR and OpenSubtitles FR into ~256 MB shards; not included here).

checkpoints/byte_tick_bptt_fresh/atom_native_step_1232000_d32_bptt_fresh.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5bb47fb1dc821ef168b0a1693db720a3105b0c7dd8cfef737ad3936aaaeac399
+size 3815607

checkpoints/byte_tick_bptt_fresh/run_metrics_1232000_d32_bptt_fresh.json

@@ -0,0 +1,234 @@
+{
+  "stream": true,
+  "data_path": null,
+  "data_glob": "data/corpus_fr_medium*.txt",
+  "shard_paths": [
+    "/workspace/exports/atom-ai/data/corpus_fr_medium.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_medium_s21.txt"
+  ],
+  "data_bytes": 1000453,
+  "data_characters": null,
+  "atomizer_version": "atomizer-v1-byte-span",
+  "feature_dim": 320,
+  "packet_count": null,
+  "train_packet_count": null,
+  "validation_packet_count": 2048,
+  "mean_packet_bytes": null,
+  "max_packet_bytes": null,
+  "min_packet_bytes": null,
+  "bytes_seen": 262144,
+  "config": {
+    "d_model": 32,
+    "n_modes": 32,
+    "n_atoms_max": 64,
+    "max_span_bytes": 1,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "learning_rate": 0.002,
+    "surface_learning_rate": 0.002,
+    "steps": 8000,
+    "start_step": 1224000,
+    "seed": 20260913,
+    "surface_alphabet": 256,
+    "field_max_rms": 3.0,
+    "energy_decay_bounds": [
+      0.3,
+      0.95
+    ],
+    "stream": true,
+    "chunk_bytes": 65536,
+    "loop_shards": true,
+    "stream_buffer": 2048,
+    "stream_prefetch": true,
+    "stream_skip_packets": 223548,
+    "last_atom_readout": true,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_rank": 64,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "enable_merge": false,
+    "payload_copy": false,
+    "merge_coherence_threshold": 0.45,
+    "merge_energy_floor": 0.08,
+    "field_loss_weight": 0.05,
+    "field_contrast_weight": 0.0,
+    "field_ignorance_weight": 0.08,
+    "field_ignorance_ablate_shared": false,
+    "field_ignorance_prompt_bank": false,
+    "field_obligatory_readout": true,
+    "field_obligatory_hard": true,
+    "printable_aux_weight": 0.0,
+    "field_next_packet_weight": 0.0,
+    "slow_every": 1,
+    "field_bptt": 64,
+    "field_leak_learnable": true
+  },
+  "packets_seen": 262144,
+  "stream_epoch": 0,
+  "train": {
+    "steps": 8000,
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "efference_steps": 0,
+    "field_bptt": 64,
+    "bptt_windows": 125,
+    "field_leak_learnable_rates": [
+      0.1651015728712082,
+      0.2223760485649109,
+      0.19814087450504303,
+      0.16405101120471954,
+      0.14837774634361267,
+      0.2066447138786316,
+      0.2014492303133011,
+      0.11219972372055054,
+      0.12348398566246033,
+      0.10896243155002594,
+      0.06129872426390648,
+      0.10106329619884491,
+      0.384555459022522,
+      0.5364075303077698,
+      0.17985180020332336,
+      0.1374603509902954,
+      0.177724689245224,
+      0.21974866092205048,
+      0.2793871760368347,
+      0.1863066852092743,
+      0.23682813346385956,
+      0.13955748081207275,
+      0.1764942854642868,
+      0.1738881915807724,
+      0.1693471372127533,
+      0.16529005765914917,
+      0.13834939897060394,
+      0.1200307309627533,
+      0.19962647557258606,
+      0.288830041885376,
+      0.20705798268318176,
+      0.23813004791736603
+    ],
+    "free_run_aux_steps": 0,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "slow_memory_loss_mean": null,
+    "slow_memory_loss_last_1000": null,
+    "start_step": 1224000,
+    "total_steps": 1232000,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "elapsed_seconds": 58.50372852999135,
+    "transitions_per_second": 136.74342133422968,
+    "first_loss": 1.4675815105438232,
+    "final_loss": 0.0031006068456918,
+    "final_total_perplexity": 1.003105418699034,
+    "finite_losses": true,
+    "finite_parameters": true,
+    "nan_or_inf": false,
+    "field_rms_before_max": 0.024285878986120224,
+    "field_rms_after_max": 0.024285878986120224,
+    "field_scale_min": 1.0,
+    "field_norm_max": 0.7764889597892761,
+    "field_limited_steps": 0,
+    "field_limited_fraction": 0.0,
+    "final_dynamics": {
+      "coupling_scale": 1.5999668836593628,
+      "energy_decay": 0.8711663484573364,
+      "phase_sync": 0.9322643876075745,
+      "energy_decay_before": 0.8711663484573364,
+      "energy_decay_repaired": false,
+      "energy_decay_bounds": [
+        0.3,
+        0.95
+      ]
+    },
+    "final_atoms_before_validation": 58,
+    "stream": true,
+    "bytes_seen": 262144
+  },
+  "validation": {
+    "transitions": 256,
+    "loss": 0.4576556962065297,
+    "total_perplexity": 1.5803647837882906,
+    "byte_loss": 0.4540158328897905,
+    "byte_perplexity": 1.5746229281089386,
+    "source": "stream_ring"
+  },
+  "checkpoint": "checkpoints/byte_tick_bptt_fresh/atom_native_step_1232000_d32_bptt_fresh.pt",
+  "reload": {
+    "n_atoms": 0,
+    "field_norm": 0.28036776185035706,
+    "time": 79.9986801147461,
+    "finite": true
+  },
+  "generation": {
+    "The future of AI is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "5061732064652070",
+      "text": "Pas de p",
+      "reload_text": "Pas de p",
+      "reload_match": true
+    },
+    "Bonjour, je m'appelle": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "5061732074726f70",
+      "text": "Pas trop",
+      "reload_text": "Pas trop",
+      "reload_match": true
+    },
+    "Valkyria Chronicles III is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "4a65207375697320",
+      "text": "Je suis ",
+      "reload_text": "Je suis ",
+      "reload_match": true
+    }
+  }
+}

checkpoints/byte_tick_bptt_fresh_k128/atom_native_step_1104000_d32_bptt_fresh_k128.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dc9d83ed4effa047c8df6e8a182a7a03bedf0f214f3e88aea5ec6503319a08b
+size 3816649

checkpoints/byte_tick_bptt_fresh_k128/run_metrics_1104000_d32_bptt_fresh_k128.json

@@ -0,0 +1,234 @@
+{
+  "stream": true,
+  "data_path": null,
+  "data_glob": "data/corpus_fr_medium*.txt",
+  "shard_paths": [
+    "/workspace/exports/atom-ai/data/corpus_fr_medium.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_medium_s21.txt"
+  ],
+  "data_bytes": 1000453,
+  "data_characters": null,
+  "atomizer_version": "atomizer-v1-byte-span",
+  "feature_dim": 320,
+  "packet_count": null,
+  "train_packet_count": null,
+  "validation_packet_count": 2048,
+  "mean_packet_bytes": null,
+  "max_packet_bytes": null,
+  "min_packet_bytes": null,
+  "bytes_seen": 131072,
+  "config": {
+    "d_model": 32,
+    "n_modes": 32,
+    "n_atoms_max": 64,
+    "max_span_bytes": 1,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "learning_rate": 0.002,
+    "surface_learning_rate": 0.002,
+    "steps": 8000,
+    "start_step": 1096000,
+    "seed": 20260913,
+    "surface_alphabet": 256,
+    "field_max_rms": 3.0,
+    "energy_decay_bounds": [
+      0.3,
+      0.95
+    ],
+    "stream": true,
+    "chunk_bytes": 65536,
+    "loop_shards": true,
+    "stream_buffer": 2048,
+    "stream_prefetch": true,
+    "stream_skip_packets": 95548,
+    "last_atom_readout": true,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_rank": 64,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "enable_merge": false,
+    "payload_copy": false,
+    "merge_coherence_threshold": 0.45,
+    "merge_energy_floor": 0.08,
+    "field_loss_weight": 0.05,
+    "field_contrast_weight": 0.0,
+    "field_ignorance_weight": 0.08,
+    "field_ignorance_ablate_shared": false,
+    "field_ignorance_prompt_bank": false,
+    "field_obligatory_readout": true,
+    "field_obligatory_hard": true,
+    "printable_aux_weight": 0.0,
+    "field_next_packet_weight": 0.0,
+    "slow_every": 1,
+    "field_bptt": 128,
+    "field_leak_learnable": true
+  },
+  "packets_seen": 131072,
+  "stream_epoch": 0,
+  "train": {
+    "steps": 8000,
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "efference_steps": 0,
+    "field_bptt": 128,
+    "bptt_windows": 63,
+    "field_leak_learnable_rates": [
+      0.13960112631320953,
+      0.20525416731834412,
+      0.12362688034772873,
+      0.22679980099201202,
+      0.19011618196964264,
+      0.2006656676530838,
+      0.18863052129745483,
+      0.13466861844062805,
+      0.1541769951581955,
+      0.11504282802343369,
+      0.07188660651445389,
+      0.08401831984519958,
+      0.1887102723121643,
+      0.38703152537345886,
+      0.20451544225215912,
+      0.21816956996917725,
+      0.20279471576213837,
+      0.23330043256282806,
+      0.20438706874847412,
+      0.21167437732219696,
+      0.19299842417240143,
+      0.16187530755996704,
+      0.14983811974525452,
+      0.17153199017047882,
+      0.1774502992630005,
+      0.15822280943393707,
+      0.23028582334518433,
+      0.18020951747894287,
+      0.19170816242694855,
+      0.21261583268642426,
+      0.1362137347459793,
+      0.20936228334903717
+    ],
+    "free_run_aux_steps": 0,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "slow_memory_loss_mean": null,
+    "slow_memory_loss_last_1000": null,
+    "start_step": 1096000,
+    "total_steps": 1104000,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "elapsed_seconds": 57.065072412005975,
+    "transitions_per_second": 140.19083235784825,
+    "first_loss": 16.306896209716797,
+    "final_loss": 0.07207974791526794,
+    "final_total_perplexity": 1.074741049004484,
+    "finite_losses": true,
+    "finite_parameters": true,
+    "nan_or_inf": false,
+    "field_rms_before_max": 0.017805438488721848,
+    "field_rms_after_max": 0.017805438488721848,
+    "field_scale_min": 1.0,
+    "field_norm_max": 0.568874716758728,
+    "field_limited_steps": 0,
+    "field_limited_fraction": 0.0,
+    "final_dynamics": {
+      "coupling_scale": 1.254072666168213,
+      "energy_decay": 0.32255834341049194,
+      "phase_sync": 0.9864845871925354,
+      "energy_decay_before": 0.32255834341049194,
+      "energy_decay_repaired": false,
+      "energy_decay_bounds": [
+        0.3,
+        0.95
+      ]
+    },
+    "final_atoms_before_validation": 58,
+    "stream": true,
+    "bytes_seen": 131072
+  },
+  "validation": {
+    "transitions": 256,
+    "loss": 0.528362570520585,
+    "total_perplexity": 1.696152703097949,
+    "byte_loss": 0.524585546465928,
+    "byte_perplexity": 1.6897583769016438,
+    "source": "stream_ring"
+  },
+  "checkpoint": "checkpoints/byte_tick_bptt_fresh_k128/atom_native_step_1104000_d32_bptt_fresh_k128.pt",
+  "reload": {
+    "n_atoms": 0,
+    "field_norm": 0.14853636920452118,
+    "time": 79.9986801147461,
+    "finite": true
+  },
+  "generation": {
+    "The future of AI is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "426f6e6e65207175",
+      "text": "Bonne qu",
+      "reload_text": "Bonne qu",
+      "reload_match": true
+    },
+    "Bonjour, je m'appelle": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "4176656320706c61",
+      "text": "Avec pla",
+      "reload_text": "Avec pla",
+      "reload_match": true
+    },
+    "Valkyria Chronicles III is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "44276163636f7264",
+      "text": "D'accord",
+      "reload_text": "D'accord",
+      "reload_match": true
+    }
+  }
+}

checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1600000_d32_bptt_k128_mix.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33e4cd75ab1bb9d7a825b32af9fca12865e88bdd2332fda4af2e983f5cf9d56b
+size 3816181

checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1680000_d32_bptt_k128_mix.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12e032796371f50e37718a2c6848c4159ffe770f9ae6e2b3c6a427f3d5a2a52b
+size 3816181

checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1720000_d32_bptt_k128_mix.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fed48094cea4f0be4d4b53dcbd5a8cf89921a7abb3844271a5fa68751f9a0b7e
+size 3816181

checkpoints/byte_tick_bptt_k128_mix/run_metrics_1600000_d32_bptt_k128_mix.json

@@ -0,0 +1,237 @@
+{
+  "stream": true,
+  "data_path": null,
+  "data_glob": "data/corpus_fr_mix_[0-9]*.txt",
+  "shard_paths": [
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_000.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_001.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_002.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_003.txt"
+  ],
+  "data_bytes": 7394592,
+  "data_characters": null,
+  "atomizer_version": "atomizer-v1-byte-span",
+  "feature_dim": 320,
+  "packet_count": null,
+  "train_packet_count": null,
+  "validation_packet_count": 2048,
+  "mean_packet_bytes": null,
+  "max_packet_bytes": null,
+  "min_packet_bytes": null,
+  "bytes_seen": 65536,
+  "config": {
+    "d_model": 32,
+    "n_modes": 32,
+    "n_atoms_max": 64,
+    "max_span_bytes": 1,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "learning_rate": 0.002,
+    "surface_learning_rate": 0.002,
+    "steps": 8000,
+    "start_step": 1592000,
+    "seed": 20260913,
+    "surface_alphabet": 256,
+    "field_max_rms": 3.0,
+    "energy_decay_bounds": [
+      0.3,
+      0.95
+    ],
+    "stream": true,
+    "chunk_bytes": 65536,
+    "loop_shards": true,
+    "stream_buffer": 2048,
+    "stream_prefetch": true,
+    "stream_skip_packets": 0,
+    "last_atom_readout": true,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_rank": 64,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "enable_merge": false,
+    "payload_copy": false,
+    "merge_coherence_threshold": 0.45,
+    "merge_energy_floor": 0.08,
+    "field_loss_weight": 0.05,
+    "field_contrast_weight": 0.0,
+    "field_ignorance_weight": 0.08,
+    "field_ignorance_ablate_shared": false,
+    "field_ignorance_prompt_bank": false,
+    "field_obligatory_readout": true,
+    "field_obligatory_hard": true,
+    "printable_aux_weight": 0.0,
+    "field_next_packet_weight": 0.0,
+    "slow_every": 1,
+    "field_bptt": 128,
+    "field_leak_learnable": true,
+    "stream_skip_bytes": 488000
+  },
+  "packets_seen": 65536,
+  "stream_epoch": 0,
+  "train": {
+    "steps": 8000,
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "efference_steps": 0,
+    "field_bptt": 128,
+    "bptt_windows": 63,
+    "field_leak_learnable_rates": [
+      0.2589353322982788,
+      0.33240318298339844,
+      0.14712077379226685,
+      0.23442696034908295,
+      0.21832816302776337,
+      0.30121657252311707,
+      0.2626681923866272,
+      0.15994516015052795,
+      0.19780592620372772,
+      0.16097243130207062,
+      0.10915376245975494,
+      0.1730736494064331,
+      0.3482528030872345,
+      0.7338636517524719,
+      0.42805999517440796,
+      0.5028843283653259,
+      0.32569214701652527,
+      0.48313575983047485,
+      0.34425604343414307,
+      0.3276020288467407,
+      0.3864366114139557,
+      0.2443476915359497,
+      0.227683424949646,
+      0.23513294756412506,
+      0.3145107328891754,
+      0.3147050440311432,
+      0.36194443702697754,
+      0.2797192633152008,
+      0.2787001132965088,
+      0.29079070687294006,
+      0.23655368387699127,
+      0.42117777466773987
+    ],
+    "free_run_aux_steps": 0,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "slow_memory_loss_mean": null,
+    "slow_memory_loss_last_1000": null,
+    "start_step": 1592000,
+    "total_steps": 1600000,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "elapsed_seconds": 75.00176022500091,
+    "transitions_per_second": 106.66416329430758,
+    "first_loss": 10.275494575500488,
+    "final_loss": 2.529386043548584,
+    "final_total_perplexity": 12.54580119607875,
+    "finite_losses": true,
+    "finite_parameters": true,
+    "nan_or_inf": false,
+    "field_rms_before_max": 0.020766643807291985,
+    "field_rms_after_max": 0.020766643807291985,
+    "field_scale_min": 1.0,
+    "field_norm_max": 0.6637617945671082,
+    "field_limited_steps": 0,
+    "field_limited_fraction": 0.0,
+    "final_dynamics": {
+      "coupling_scale": 1.3370531797409058,
+      "energy_decay": 0.3041408360004425,
+      "phase_sync": 0.9343012571334839,
+      "energy_decay_before": 0.3041408360004425,
+      "energy_decay_repaired": false,
+      "energy_decay_bounds": [
+        0.3,
+        0.95
+      ]
+    },
+    "final_atoms_before_validation": 58,
+    "stream": true,
+    "bytes_seen": 65536
+  },
+  "validation": {
+    "transitions": 256,
+    "loss": 1.9196898239679285,
+    "total_perplexity": 6.818843099543278,
+    "byte_loss": 1.9157627062716642,
+    "byte_perplexity": 6.792117212339081,
+    "source": "stream_ring"
+  },
+  "checkpoint": "checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1600000_d32_bptt_k128_mix.pt",
+  "reload": {
+    "n_atoms": 0,
+    "field_norm": 0.19758206605911255,
+    "time": 79.9986801147461,
+    "finite": true
+  },
+  "generation": {
+    "The future of AI is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "426f6e6e65207175",
+      "text": "Bonne qu",
+      "reload_text": "Bonne qu",
+      "reload_match": true
+    },
+    "Bonjour, je m'appelle": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "426f6e6e65207175",
+      "text": "Bonne qu",
+      "reload_text": "Bonne qu",
+      "reload_match": true
+    },
+    "Valkyria Chronicles III is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "4176656320706c61",
+      "text": "Avec pla",
+      "reload_text": "Avec pla",
+      "reload_match": true
+    }
+  }
+}

checkpoints/byte_tick_bptt_k128_mix/run_metrics_1680000_d32_bptt_k128_mix.json

@@ -0,0 +1,237 @@
+{
+  "stream": true,
+  "data_path": null,
+  "data_glob": "data/corpus_fr_mix_[0-9]*.txt",
+  "shard_paths": [
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_000.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_001.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_002.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_003.txt"
+  ],
+  "data_bytes": 7394592,
+  "data_characters": null,
+  "atomizer_version": "atomizer-v1-byte-span",
+  "feature_dim": 320,
+  "packet_count": null,
+  "train_packet_count": null,
+  "validation_packet_count": 2048,
+  "mean_packet_bytes": null,
+  "max_packet_bytes": null,
+  "min_packet_bytes": null,
+  "bytes_seen": 65536,
+  "config": {
+    "d_model": 32,
+    "n_modes": 32,
+    "n_atoms_max": 64,
+    "max_span_bytes": 1,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "learning_rate": 0.002,
+    "surface_learning_rate": 0.002,
+    "steps": 8000,
+    "start_step": 1672000,
+    "seed": 20260913,
+    "surface_alphabet": 256,
+    "field_max_rms": 3.0,
+    "energy_decay_bounds": [
+      0.3,
+      0.95
+    ],
+    "stream": true,
+    "chunk_bytes": 65536,
+    "loop_shards": true,
+    "stream_buffer": 2048,
+    "stream_prefetch": true,
+    "stream_skip_packets": 0,
+    "last_atom_readout": true,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_rank": 64,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "enable_merge": false,
+    "payload_copy": false,
+    "merge_coherence_threshold": 0.45,
+    "merge_energy_floor": 0.08,
+    "field_loss_weight": 0.05,
+    "field_contrast_weight": 0.0,
+    "field_ignorance_weight": 0.08,
+    "field_ignorance_ablate_shared": false,
+    "field_ignorance_prompt_bank": false,
+    "field_obligatory_readout": true,
+    "field_obligatory_hard": true,
+    "printable_aux_weight": 0.0,
+    "field_next_packet_weight": 0.0,
+    "slow_every": 1,
+    "field_bptt": 128,
+    "field_leak_learnable": true,
+    "stream_skip_bytes": 568000
+  },
+  "packets_seen": 65536,
+  "stream_epoch": 0,
+  "train": {
+    "steps": 8000,
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "efference_steps": 0,
+    "field_bptt": 128,
+    "bptt_windows": 63,
+    "field_leak_learnable_rates": [
+      0.274522066116333,
+      0.3533155024051666,
+      0.15473133325576782,
+      0.2395753115415573,
+      0.22116175293922424,
+      0.31720712780952454,
+      0.27202317118644714,
+      0.16598258912563324,
+      0.20310036838054657,
+      0.16193081438541412,
+      0.10908772796392441,
+      0.18383222818374634,
+      0.35520121455192566,
+      0.75541752576828,
+      0.43806663155555725,
+      0.529101550579071,
+      0.33727988600730896,
+      0.5073504447937012,
+      0.3601907193660736,
+      0.3478285074234009,
+      0.41036245226860046,
+      0.2591860592365265,
+      0.2394794225692749,
+      0.24186107516288757,
+      0.33263686299324036,
+      0.3232515752315521,
+      0.37150341272354126,
+      0.2870956361293793,
+      0.28565606474876404,
+      0.30195820331573486,
+      0.24409471452236176,
+      0.4355026185512543
+    ],
+    "free_run_aux_steps": 0,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "slow_memory_loss_mean": null,
+    "slow_memory_loss_last_1000": null,
+    "start_step": 1672000,
+    "total_steps": 1680000,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "elapsed_seconds": 79.34130910399836,
+    "transitions_per_second": 100.83019917800732,
+    "first_loss": 3.9205265045166016,
+    "final_loss": 0.005072177387773991,
+    "final_total_perplexity": 1.005085062655744,
+    "finite_losses": true,
+    "finite_parameters": true,
+    "nan_or_inf": false,
+    "field_rms_before_max": 0.016347821801900864,
+    "field_rms_after_max": 0.016347821801900864,
+    "field_scale_min": 1.0,
+    "field_norm_max": 0.5221506357192993,
+    "field_limited_steps": 0,
+    "field_limited_fraction": 0.0,
+    "final_dynamics": {
+      "coupling_scale": 1.3326725959777832,
+      "energy_decay": 0.30556628108024597,
+      "phase_sync": 0.8960594534873962,
+      "energy_decay_before": 0.30556628108024597,
+      "energy_decay_repaired": false,
+      "energy_decay_bounds": [
+        0.3,
+        0.95
+      ]
+    },
+    "final_atoms_before_validation": 58,
+    "stream": true,
+    "bytes_seen": 65536
+  },
+  "validation": {
+    "transitions": 256,
+    "loss": 1.1553669280856411,
+    "total_perplexity": 3.175188269559367,
+    "byte_loss": 1.1512991201693827,
+    "byte_perplexity": 3.1622984480139547,
+    "source": "stream_ring"
+  },
+  "checkpoint": "checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1680000_d32_bptt_k128_mix.pt",
+  "reload": {
+    "n_atoms": 0,
+    "field_norm": 0.1693684607744217,
+    "time": 79.9986801147461,
+    "finite": true
+  },
+  "generation": {
+    "The future of AI is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "4a65207375206427",
+      "text": "Je su d'",
+      "reload_text": "Je su d'",
+      "reload_match": true
+    },
+    "Bonjour, je m'appelle": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "4a65207375206427",
+      "text": "Je su d'",
+      "reload_text": "Je su d'",
+      "reload_match": true
+    },
+    "Valkyria Chronicles III is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "4a65207375206427",
+      "text": "Je su d'",
+      "reload_text": "Je su d'",
+      "reload_match": true
+    }
+  }
+}

checkpoints/byte_tick_bptt_k128_mix/run_metrics_1720000_d32_bptt_k128_mix.json

@@ -0,0 +1,237 @@
+{
+  "stream": true,
+  "data_path": null,
+  "data_glob": "data/corpus_fr_mix_[0-9]*.txt",
+  "shard_paths": [
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_000.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_001.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_002.txt",
+    "/workspace/exports/atom-ai/data/corpus_fr_mix_003.txt"
+  ],
+  "data_bytes": 7394592,
+  "data_characters": null,
+  "atomizer_version": "atomizer-v1-byte-span",
+  "feature_dim": 320,
+  "packet_count": null,
+  "train_packet_count": null,
+  "validation_packet_count": 2048,
+  "mean_packet_bytes": null,
+  "max_packet_bytes": null,
+  "min_packet_bytes": null,
+  "bytes_seen": 65536,
+  "config": {
+    "d_model": 32,
+    "n_modes": 32,
+    "n_atoms_max": 64,
+    "max_span_bytes": 1,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "learning_rate": 0.002,
+    "surface_learning_rate": 0.002,
+    "steps": 8000,
+    "start_step": 1712000,
+    "seed": 20260913,
+    "surface_alphabet": 256,
+    "field_max_rms": 3.0,
+    "energy_decay_bounds": [
+      0.3,
+      0.95
+    ],
+    "stream": true,
+    "chunk_bytes": 65536,
+    "loop_shards": true,
+    "stream_buffer": 2048,
+    "stream_prefetch": true,
+    "stream_skip_packets": 0,
+    "last_atom_readout": true,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_rank": 64,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "enable_merge": false,
+    "payload_copy": false,
+    "merge_coherence_threshold": 0.45,
+    "merge_energy_floor": 0.08,
+    "field_loss_weight": 0.05,
+    "field_contrast_weight": 0.0,
+    "field_ignorance_weight": 0.08,
+    "field_ignorance_ablate_shared": false,
+    "field_ignorance_prompt_bank": false,
+    "field_obligatory_readout": true,
+    "field_obligatory_hard": true,
+    "printable_aux_weight": 0.0,
+    "field_next_packet_weight": 0.0,
+    "slow_every": 1,
+    "field_bptt": 128,
+    "field_leak_learnable": true,
+    "stream_skip_bytes": 608000
+  },
+  "packets_seen": 65536,
+  "stream_epoch": 0,
+  "train": {
+    "steps": 8000,
+    "efference_every": 0,
+    "free_run_aux_every": 0,
+    "free_run_aux_horizon": 4,
+    "free_run_aux_weight": 1.0,
+    "efference_steps": 0,
+    "field_bptt": 128,
+    "bptt_windows": 63,
+    "field_leak_learnable_rates": [
+      0.2812689244747162,
+      0.3619617223739624,
+      0.15832066535949707,
+      0.2400067150592804,
+      0.2243746519088745,
+      0.318925678730011,
+      0.272644966840744,
+      0.16585779190063477,
+      0.20562595129013062,
+      0.16399800777435303,
+      0.1094951331615448,
+      0.18695788085460663,
+      0.35893818736076355,
+      0.7612906098365784,
+      0.4403144121170044,
+      0.5398162007331848,
+      0.3397994935512543,
+      0.5101497769355774,
+      0.37088844180107117,
+      0.36206647753715515,
+      0.4253099262714386,
+      0.26382380723953247,
+      0.23801271617412567,
+      0.2422337830066681,
+      0.3436702787876129,
+      0.32509443163871765,
+      0.3688523471355438,
+      0.2930602729320526,
+      0.2839611768722534,
+      0.3078254163265228,
+      0.24626755714416504,
+      0.4446651041507721
+    ],
+    "free_run_aux_steps": 0,
+    "last_atom_scale": 1.0,
+    "last_atom_scale_adaptive": false,
+    "last_atom_scale_min": 0.0,
+    "last_atom_scale_max": 1.0,
+    "field_hard_couple_scale": 0.0,
+    "last_atom_dropout": 0.5,
+    "field_leak": 0.05,
+    "field_leak_spectrum": null,
+    "field_leak_spectrum_gain": "none",
+    "field_leak_spectrum_slow_frac": 1.0,
+    "slow_mode_readout": false,
+    "slow_readout_ahat": "exclude",
+    "slow_readout_lr": null,
+    "slow_memory_loss": 0.0,
+    "slow_memory_lags": [
+      16,
+      32,
+      64
+    ],
+    "slow_memory_target": "bag",
+    "slow_memory_loss_mean": null,
+    "slow_memory_loss_last_1000": null,
+    "start_step": 1712000,
+    "total_steps": 1720000,
+    "episode_length": 8000,
+    "episode_reset": true,
+    "atom_flush_every": 64,
+    "elapsed_seconds": 74.49479142199561,
+    "transitions_per_second": 107.39005838249639,
+    "first_loss": 15.9537992477417,
+    "final_loss": 1.1094855070114136,
+    "final_total_perplexity": 3.032797639849954,
+    "finite_losses": true,
+    "finite_parameters": true,
+    "nan_or_inf": false,
+    "field_rms_before_max": 0.021175356581807137,
+    "field_rms_after_max": 0.021175356581807137,
+    "field_scale_min": 1.0,
+    "field_norm_max": 0.6768553256988525,
+    "field_limited_steps": 0,
+    "field_limited_fraction": 0.0,
+    "final_dynamics": {
+      "coupling_scale": 1.3405709266662598,
+      "energy_decay": 0.3069625794887543,
+      "phase_sync": 0.889988124370575,
+      "energy_decay_before": 0.3069625794887543,
+      "energy_decay_repaired": false,
+      "energy_decay_bounds": [
+        0.3,
+        0.95
+      ]
+    },
+    "final_atoms_before_validation": 58,
+    "stream": true,
+    "bytes_seen": 65536
+  },
+  "validation": {
+    "transitions": 256,
+    "loss": 2.015277966503163,
+    "total_perplexity": 7.5028126219644085,
+    "byte_loss": 2.0113292363094653,
+    "byte_perplexity": 7.473244456001308,
+    "source": "stream_ring"
+  },
+  "checkpoint": "checkpoints/byte_tick_bptt_k128_mix/atom_native_step_1720000_d32_bptt_k128_mix.pt",
+  "reload": {
+    "n_atoms": 0,
+    "field_norm": 0.20565296709537506,
+    "time": 79.9986801147461,
+    "finite": true
+  },
+  "generation": {
+    "The future of AI is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "4f75692c20746f75",
+      "text": "Oui, tou",
+      "reload_text": "Oui, tou",
+      "reload_match": true
+    },
+    "Bonjour, je m'appelle": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "41766563206c6169",
+      "text": "Avec lai",
+      "reload_text": "Avec lai",
+      "reload_match": true
+    },
+    "Valkyria Chronicles III is": {
+      "temperature": 0.8,
+      "top_k": 5,
+      "max_length": 20,
+      "bytes_hex": "41766563206c6169",
+      "text": "Avec lai",
+      "reload_text": "Avec lai",
+      "reload_match": true
+    }
+  }
+}

checkpoints/gpu_smoke_b512/atom_native_step_4000_smoke_b512.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d5fb5e1bd86cea3650f61e5678558401d527b434cf50e335e87a7774dc8f7c6f
+size 10096441

checkpoints/gpu_smoke_b512/progress.tsv

@@ -0,0 +1,5 @@
+time	tick	bytes_total	bytes_per_s	train_loss	byte_ce	persist	grad_norm	leak_mean	val_ce	val_field_alone	ckpt	samples
+2026-09-30T21:05:11	1000	512000	39045	1.8098	1.8059	0.0780	0.170	0.3053	1.2957	1.3174	checkpoints/smoke_b512/atom_native_step_1000_smoke_b512.pt	{"Bonjour, comment ça va ?": "Je d'escord une cheauverte, que la cheur quellantre cairente ces puirent à tourariont li ren l", "Qui es-tu ?": "Avec toute set mégérat.\nUtilisateur: Tu es de claisent à lis de, d'un les prordeveux, don de", "Il était une fois": "Avec toute set mégérat.\nUtilisateur: Tu es de claisent à lis d'hamant, pomme à sons.\n\n"}
+2026-09-30T21:05:31	2000	1024000	37840	1.7619	1.7580	0.0781	0.116	0.3064	1.2615	1.2934	checkpoints/smoke_b512/atom_native_step_2000_smoke_b512.pt	{"Bonjour, comment ça va ?": "Oui, que pelle ent�. Des, sa chala posse autres pé supitais couter etre pes, de tromme l'il po", "Qui es-tu ?": "Oui, que d'aimons epps l'irce qu'il dempies pris étatre qui avec toure ce phourcurées, dans s", "Il était une fois": "Oui, que d'ai avoun, je perle vais de consire que actraient léchi du de tes d'accord ?\nC'amait"}
+2026-09-30T21:05:51	3000	1536000	38582	1.7347	1.7308	0.0773	0.117	0.3073	1.2457	1.2780	checkpoints/smoke_b512/atom_native_step_3000_smoke_b512.pt	{"Bonjour, comment ça va ?": "Je susi d'accord une sestent à de le conss les pématait quet les. Arton, cron véfli, a fait ", "Qui es-tu ?": "Avec tout precrertevec loier de contaient les de nois.\nAssistant: D'accord je t'éce quide le t", "Il était une fois": "Avec toins tapoit agacler épours de que d'acce que sorent deur l'ondes talent plus prosanient "}
+2026-09-30T21:06:11	4000	2048000	40270	1.7356	1.7317	0.0778	0.098	0.3086	1.2307	1.2644	checkpoints/smoke_b512/atom_native_step_4000_smoke_b512.pt	{"Bonjour, comment ça va ?": "Je susi désabonteturs dans par sans est un pas se boit cembluienté poure léchommentèmes de ", "Qui es-tu ?": "Je susi du me machame. Carle que les avaisseur avait elles plus d'inest sersur alis, checle de ", "Il était une fois": "Je susi désa mouve. l'ames a faisse n'entres le faiser suse ses lisser este, qui et lail fait "}