| { |
| "policy": "educational.policy", |
| "version": 1, |
| "subject": "duplicate a successful mindXtrain run", |
| "exemplar": { |
| "generation": 39, |
| "model": "PYTHAI/mindXtrain39", |
| "why_this_one": "the newest generation the imprint gate accepted; of the 37 attempts logged since, 30 proof_rejected (41, 46-74), 6 train_failed (40, 42-45, 75), none accepted" |
| }, |
| "claim": "A 135M model, two CPU cores and 70 minutes are enough to move proof-of-recall by +0.10. That is the whole of the claim: recall of a corpus, not identity and not reasoning.", |
| "measured": { |
| "_source": "gen39's own train.log (published at PYTHAI/mindXtrain39/train.log) and the ascent log entry for generation 39 — measured, not reconstructed", |
| "steps": 116, |
| "epochs": 2, |
| "train_runtime_s": 4201, |
| "ascent_wall_s": 4221.5, |
| "train_loss": 1.65, |
| "eval_loss": 1.225, |
| "eval_entropy": 1.445, |
| "eval_tokens": 348500, |
| "train_examples_tokenized": 460, |
| "s_per_step_mean": 36.2, |
| "imprint": { |
| "delta_recall": 0.1002, |
| "imprinted": true, |
| "stage": "accepted", |
| "gate": "mindXtrain imprint (proof of recall)" |
| } |
| }, |
| "recipe": { |
| "_source": "the run recipe shape mindX writes per ascent (data/godel/ascend/genN/run.yaml); values here are the ones gen39's log confirms", |
| "model": { |
| "name": "HuggingFaceTB/SmolLM2-135M", |
| "attn_implementation": "eager", |
| "torch_dtype": "float32" |
| }, |
| "data": { |
| "source": "mindx_dreams", |
| "path": "data/memory/curated", |
| "seq_len": 1024, |
| "packing": true, |
| "eval_split": 0.1, |
| "include_evolutions": true, |
| "max_samples": 1024 |
| }, |
| "train": { |
| "backend": "trl_cpu", |
| "method": { |
| "kind": "lora", |
| "r": 16, |
| "alpha": 32, |
| "dropout": 0.0, |
| "target_modules": [ |
| "q_proj", |
| "k_proj", |
| "v_proj", |
| "o_proj" |
| ] |
| }, |
| "optimizer": { |
| "name": "adamw_torch", |
| "lr": 0.0001 |
| }, |
| "schedule": { |
| "type": "cosine", |
| "warmup_ratio": 0.03, |
| "epochs": 2 |
| }, |
| "batch": { |
| "per_device": 1, |
| "grad_accum": 8 |
| }, |
| "precision": "float32", |
| "cpu_throttle": { |
| "percent": 33, |
| "nice": 19 |
| } |
| }, |
| "hardware_measured_on": { |
| "cpu_cores": 2, |
| "cpu_model": "AMD EPYC 7543P", |
| "ram_gb": 7.8, |
| "gpu": null |
| } |
| }, |
| "duplicate": { |
| "framework": "https://github.com/professor-codephreak/mindXtrain", |
| "steps": [ |
| "uv sync --extra ml # trl + transformers + peft + accelerate", |
| "mindxtrain init -t mindx_fallback_qwen3_1_5b_cpu_smoke -o run.yaml", |
| "edit run.yaml to the recipe below (LoRA r16/α32 on q,k,v,o · lr 1e-4 cosine · 2 epochs · packing · seq 1024 · eval_split 0.1)", |
| "mindxtrain train run.yaml --out out/runs --cpu-percent 33 --cpu-nice 19", |
| "mindxtrain imprint --config run.yaml # the gate: recall BEFORE vs AFTER", |
| "keep the run only if delta > the calibrated floor; otherwise it is a rejected generation and is recorded as one", |
| "mindxtrain serve --config run.yaml --to ollama --tag <name> # only after a positive imprint" |
| ], |
| "corpus": { |
| "what": "mindX's curated machine.dream corpus (prose in the first person + Gödel decisions + persona rows)", |
| "hub": "https://huggingface.co/datasets/PYTHAI/mindXascension/tree/main/machine.dream", |
| "rule": "the corpus is rebuilt before every ascent; provenance per row in PROVENANCE.jsonl" |
| } |
| }, |
| "gate": { |
| "name": "imprint", |
| "metric": "token-Jaccard recall of the corpus voice, after minus before", |
| "floor": "calibrated by scripts/calibrate_min_delta.py (an untrained random-init adapter is the null); provisional 0.02", |
| "decoding": { |
| "do_sample": false, |
| "repetition_penalty": 1.3, |
| "no_repeat_ngram_size": 3 |
| }, |
| "honesty": "a positive imprint proves recall. It does not prove identity: the coach measured 16% identity on this lineage." |
| }, |
| "for_the_coach": { |
| "read": [ |
| "/insight/hf/coach", |
| "/insight/hf/coach/results", |
| "/insight/godel/ascend" |
| ], |
| "act": [ |
| "POST /hf/spar/auto (score a generation)", |
| "POST /hf/coach/recommend", |
| "POST /hf/coach/recipe (adopt)" |
| ], |
| "rule": "the coach refuses a rung the ladder evidence already rejected" |
| }, |
| "orchestration": { |
| "mastermind": "https://mastermind.pythai.net", |
| "node": "https://mindx.pythai.net", |
| "note": "Mastermind is the strategic layer that decides a campaign is worth running; mindXtrain is what runs it." |
| } |
| } |
|
|
