vynly/mica-flame-w-0.3.2

MICA Flame-W 0.3.2

A newer release is available: Flame-W0.3.5 includes a verified CPU notebook and improved held-out text prediction. This historical checkpoint remains available for reproducibility. Sentence quality and broad reasoning remain experimental.

Flame-W predicts English word symbols with a Minimal Inference Cellular Automaton. The same B-740 checkpoint from Flame-W 0.3.1 runs learned integer rewrite rules on a ring of cells. This update adds a second integer memory readout after 64 fed symbols and uses a lighter anti-generic penalty when decoding sentences from long prompts. There is no transformer, GRU or retrieval generator in inference.

The model is experimental. Many raw sentences are still incoherent, especially on long or noisy text. The sentence comparison below is small and rated by one person; it is evidence of a limited improvement, not general conversational ability.

What changed

File or component	0.3.2 behavior
model.mica	Unchanged B-740 integer cellular automaton, SHA-256 1f4d550386930c8485534368033532301f2db9a49a14a69fd35b702d9e1a1f6d
memory.npz	Unchanged integer recall and association bank through 64 fed word symbols, SHA-256 5fda5a6c744749932bfba7f18352ce532405e60fca7d64fd99b705c28b928be7
memory_long.npz	New integer readout bank after 64 symbols, SHA-256 f2abb7913dd626fb75cfeb96dd97c17c7d7ecbaed88f54c95fdb986734469f3b
Sentence decoder	Same beam search and constraints; BOS anti-generic coefficient 0.5 through 64 prompt symbols, 0.25 for longer prompts

The checkpoint and two memories total 35,458,916 bytes (35.46 MB decimal). The extra memory bank is 673,650 bytes. Word-symbol vocabulary and inference geometry are unchanged. All cellular rules, scores and memory bonuses remain integer operations; the decoder uses normalized log probabilities to rank completed candidates as before.

Measured results

Exact exported integer engine, bits per word symbol plus EOS from BOS, lower is better:

Set	Published 0.3.1	0.3.2	Paired 95% CI for change
Short clean500, 7,320 targets	5.208428	5.208428	Exactly unchanged
Long clean500, 100,618 targets	8.305251	8.082371	-0.222880 [-0.241879, -0.203769]

A new disjoint 128-record development block confirmed the long-context improvement: 8.580365→8.329670 bits/target, paired change -0.250696 [-0.285911, -0.213699]. The short route is identical. These clean loss sets were inspected during earlier research, so treat the new development block and the blind sentence block below as the new decision evidence. The long development block was held out from fitting the new memory bank but came from the original cellular model's training split; the clean long500 is a separate record set.

On 20 new long development prompts, a blinded single rater preferred the 0.3.2 continuation 4 times and 0.3.1 zero times, with 16 ties. On 20 previously unused long clean prompts, it preferred 0.3.2 five times, 0.3.1 zero times, with 15 ties and zero additional severe failures. The short-prompt decoder and model route are unchanged. The fixed decoder change and gate were written before the clean prompts were generated. This sample is too small for a broad quality claim, and 15 of 20 clean pairs remained unusable under the rating rubric.

Two unedited examples from that clean comparison:

• Prompt ending when you find yourself with a few → 0.3.1: different foods that you can do it.; 0.3.2: different types of foods.
• Prompt ending the brightest comet during the first half of 2013 and → 0.3.2: comet flew closer, comet fell comet. This shows the remaining repetition problem.

Tiny Theory of Mind and the five leaderboard tasks have not been rerun for this release; the 0.3.1 results must not be presented as 0.3.2 scores. run_tom.py is provided to reproduce a future measurement of the new dual-memory route.

Use

pip install numpy
python generate.py "I owe my life to you"
python generate.py --mode suggest "Can you send me"

generate.py loads model.mica, both memories and vocab.json from this folder. --no-memory runs the original automaton without either bank. The model is deterministic on the same prompt and mode.

Source code and evaluation reports are in the MICA project. The MICA model collection lists the released family. The bundled LICENSE contains the non-commercial terms.