{
 "study": "what-a-slot-buys part 2: better, or only cheaper?",
 "written": "2026-10-30",
 "author": "ccr-fc751e34-gmn18l",
 "seed_base": 20261029,
 "arms": {
  "V": {
   "tokenizer": "../whose-words-arrive-whole/tok/files/meta-llama3/tokenizer.json",
   "sha256": "e134af98b985517b4f068e3755ae90d4e9cd2d45d328325dc503f1c6b2d06cc7",
   "training_tokens": 13847391,
   "steps": 3380
  },
  "R100": {
   "tokenizer": "out/llama3/K100.tokenizer.json",
   "sha256": "f687a257e9cc574c6a72fc32fb86ffbce8083e1cfc307dc5bbd9f8a7dff88bb1",
   "training_tokens": 11026543,
   "steps": 2692
  },
  "R749": {
   "tokenizer": "out/llama3/K1000.tokenizer.json",
   "sha256": "eb0bd1c59d26dd63300418cd5f1f9f96ffb987b987c694a794bfcf19aa698e2a",
   "training_tokens": 8338164,
   "steps": 2035
  }
 },
 "languages": {
  "heavy": [
   "shn_Mymr",
   "sat_Olck",
   "ory_Orya",
   "dzo_Tibt",
   "mya_Mymr",
   "bod_Tibt",
   "tzm_Tfng",
   "kat_Geor",
   "mal_Mlym",
   "kan_Knda",
   "lao_Laoo",
   "khm_Khmr",
   "hye_Armn",
   "sin_Sinh",
   "tel_Telu",
   "pan_Guru"
  ],
  "moderate": [
   "ckb_Arab",
   "bak_Cyrl",
   "heb_Hebr",
   "mar_Deva"
  ],
  "controls": [
   "eng_Latn",
   "nld_Latn",
   "deu_Latn",
   "ast_Latn"
  ]
 },
 "training_text": {
  "source": "lib/lmdata.py LM-TRAIN (FineWeb-2 / FineWeb draws minus MERGE, VALUE, WEB-TEST)",
  "bytes_per_language": 1000000,
  "documents": "ascending h, whole documents, each encoded alone after <|begin_of_text|>; the same documents in every arm"
 },
 "model": {
  "layers": 6,
  "width": 256,
  "heads": 4,
  "context": 512,
  "tied_embeddings": true,
  "vocabulary_rows": 128256
 },
 "training": {
  "optimizer": "AdamW (0.9, 0.95), weight decay 0.1",
  "peak_lr": 0.002,
  "warmup": "2% of steps",
  "schedule": "cosine to 0.1 of peak",
  "batch": "8 x 512 tokens",
  "passes": 1,
  "regime": "equal bytes",
  "seeds": [
   0,
   1,
   2
  ],
  "code": "part2/lm.py"
 },
 "measure": {
  "test": "FLORES-200 devtest, each sentence alone after <|begin_of_text|>",
  "bpb": "sum of sentence bits / sum of sentence UTF-8 bytes per language",
  "delta": "mean over seeds of BPB_R(L) - BPB_V(L), seeds paired",
  "dose": "r(L) = 1 - N_R749(L)/N_V(L), FLORES devtest tokens, part 1"
 },
 "pilot": {
  "models": "V, seeds 0 and 1, FLORES dev",
  "per_language_sd_pooled": 0.0118,
  "heavy_mean_sd": 0.0058,
  "controls_mean_sd": 0.0095
 },
 "predictions": {
  "Q1": {
   "arm": "R749",
   "group": "heavy",
   "mean_delta_below": -0.01,
   "and": "95% sentence-bootstrap interval of the group mean delta excludes zero"
  },
  "Q2": {
   "arm": "R749",
   "group": "controls",
   "mean_delta_at_most": 0.016
  },
  "Q3": {
   "arm": "R749",
   "languages": "heavy + moderate (20)",
   "spearman_delta_vs_dose_at_most": -0.4
  },
  "Q4": {
   "group": "heavy",
   "mean_delta_R749_at_most_mean_delta_R100": true
  }
 },
 "reported_not_tested": [
  "per-language deltas with intervals",
  "R100's Q1 and Q2 analogues",
  "steps, wall-clock and training curves per arm",
  "WEB-TEST bits per byte (first 200 KB per language, 512-token windows with 256 of context)"
 ]
}