# fx2-ft48-q256-b7-os-r5 (potatoheadjones)

**14145194 B** = `archive` 11883298 + `decompressor.zip` 2261896. Method: fx2-cmix-transformer (floating-bear's enwik8 port of upstream 83f2603)
with the transformer weights replaced by checkpoint `models/ft48-q256-b7-6m-q4-fp32.tch` (7 quantization levels); decoder-only bundle (`FX2_DECODER_ONLY`, `pack_bundle.py --decoder`).

| file | bytes | sha256 |
|---|---:|---|
| archive | 11883298 | ed07695c9969da7b8bd0ef96fd456f1f2e4b9349cb9ea3d6757619a5d0ea8346 |
| decompressor.zip | 2261896 | 1a80df38e0b3163beeb1df06926fd43f2905ff567dd0a15421a758e1829d1e31 |

Decode: `unzip decompressor.zip -d dec && python3 dec/d archive enwik8.out` on Ubuntu 24.04 (glibc 2.39), x86-64-v3, one core, ~6 GB RAM.
Build: clang-17/lld-17, `PGO=0 UPX=0 ./build_and_construct_comp.sh`, -march=x86-64-v3 -mrecip=none; AVX2-only and no-reciprocal checks passed.
Build-time cmix knobs (`scripts/fx2_knobs.py`; stock = 923/3000/0.03/1.0): SEED=923 UPDATE_LIMIT=3000 FX2_LSTM_LR=0.03 FX2_MIX_LR_SCALE=1.0. MIXER CLAMP: mixer-input probability clamp FX2_MIX_CLAMP=1.0e-5 instead of stock 1.0e-4 (swagged-out-bot's measurement; encoder and decoder share it). WEIGHT CODING: the weights member is re-coded with swagged-out-bot's ENC_INT3 (3-bit width-matched tree + count-adaptive rate, cap 7; their result 20260919-191036-788) and the decoder carries their loader patch; credit for that saving: swagged-out-bot. SIDE TENSORS: the float32 side tensors (convolutions, gates, biases) are rounded to 5 mantissa bits before coding (scripts/fx2_round_f32.py); still float32 on disk, decoder unchanged. OPT LEVEL: built with FX2_OPT=-Os for the model code and the transformer objects (encoder and decoder share it; a -O3 build cannot decode this archive).
Round trip: see logs/verify.log (`tools/verify_bundle.sh`, fresh directory, cmp vs canonical enwik8).
Transformer loss during compression: 1.176751 nats/token (full stream).
