betterwithage commited on
Commit
b1332c2
·
verified ·
1 Parent(s): 7810a96

restore: recover the pre-batch README from git history; verified-links section appended

Browse files
Files changed (1) hide show
  1. README.md +118 -21
README.md CHANGED
@@ -1,34 +1,134 @@
1
  ---
2
  license: apache-2.0
3
- library_name: transformers
4
  tags:
5
- - szl-holdings
6
- - receipts-backed
 
 
 
 
 
 
 
 
 
 
 
7
  ---
8
 
9
- # Szl Training Scripts
 
 
 
 
 
 
 
10
 
11
- > Szl Training Scripts — part of the SZL Holdings AI estate
 
 
 
12
 
13
- [![License](https://img.shields.io/badge/license-Apache--2.0-22d3ee.svg)](https://github.com/szl-holdings)
14
- [![Receipts](https://img.shields.io/badge/receipts-on__chain-34d399.svg)](https://huggingface.co/datasets/SZLHOLDINGS/frontier-54-runs)
15
- [![Source](https://img.shields.io/badge/source-GitHub-64748b.svg)](https://github.com/szl-holdings)
16
 
 
17
 
18
- ## What this is
19
 
20
- This model is part of the SZL Holdings estate: an AI research and product portfolio where every claim carries a receipt. Measured results for this artifact live on the public receipts chain — we publish numbers only where a signed receipt backs them.
21
 
22
- ## Provenance
23
 
24
- - **Source:** [szl-holdings on GitHub](https://github.com/szl-holdings)
25
- - **Receipts chain:** [SZLHOLDINGS/frontier-54-runs](https://huggingface.co/datasets/SZLHOLDINGS/frontier-54-runs) — every claim on this card traces to a hash-chained receipt
26
- - **License:** Apache-2.0
27
 
 
28
 
29
- ## Why it matters (for investors)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
30
 
31
- SZL Holdings builds receipt-backed AI: measured honesty, a machine-checked constitution, and a public hash-chained evidence chain. This artifact is one piece of that estate.
32
 
33
  ## The estate
34
 
@@ -36,9 +136,6 @@ SZL Holdings builds receipt-backed AI: measured honesty, a machine-checked const
36
  |---|---|
37
  | Company | [szlholdings.com](https://szlholdings.com) |
38
  | Sponsors | [Become a sponsor](https://szlholdings.com/sponsors.html) |
 
 
39
  | Live dashboard | [SI command center](https://szlholdings-szlholdings-si.hf.space) |
40
- | Living anatomy | [Anatomy](https://szlholdings-anatomy.hf.space) |
41
- | Autonomy ladder | [The 0.3% partition](https://github.com/szl-holdings/szl-gpu-bridge/blob/main/docs/AUTONOMY_LADDER.md) |
42
- | Training curriculum | [Ouroboros](https://github.com/szl-holdings/szl-gpu-bridge/blob/main/docs/OUROBOROS_CURRICULUM.md) |
43
-
44
- *This card is generated by [tools/hf_cards](https://github.com/szl-holdings/szl-gpu-bridge/tree/main/tools/hf_cards) — links verified at build time.*
 
1
  ---
2
  license: apache-2.0
3
+ library_name: other
4
  tags:
5
+ - governed-ai
6
+ - szl-holdings
7
+ - doctrine-v11
8
+ - estate
9
+ - hub
10
+ - training-scripts
11
+ szl:
12
+ doctrine: v11-LOCKED
13
+ artifact_class: SCRIPTS_AND_RECEIPTS
14
+ weights: UNAVAILABLE
15
+ evals: NOT_APPLICABLE
16
+ publication_eligible: false
17
+ card_resync: 2026-09-25
18
  ---
19
 
20
+ <!-- SZL-ARTIFACT-CLASS:CODE_SPEC_MIRROR -->
21
+ > [!IMPORTANT]
22
+ > **This repository is a code/spec mirror — NOT a model.** It contains no model
23
+ > weights, is not trainable, and will never produce inference. It is retained in
24
+ > the models namespace for historical continuity only. The canonical source of
25
+ > truth is the corresponding repository in the [szl-holdings GitHub org]
26
+ > (https://github.com/szl-holdings). For actual estate models, see the
27
+ > [SZLHOLDINGS trained fleet](https://huggingface.co/models?other=trained-model&author=SZLHOLDINGS).
28
 
29
+ > **CARD honesty (SZL cookie):** This Hub entry is a **card / doctrine / kernel / witness** surface.
30
+ > It does **not** claim production neural weights unless a MEASURED train receipt is linked below.
31
+ > States used: `NOT_TRAINED` · `CARD_ONLY` · `MEASURED` · `UNAVAILABLE` · `NOT_CLAIMED`.
32
+ > Production promote requires Stephen `GO WEIGHT <model-id>`. Soft Shield: Control before action. Evidence after.
33
 
34
+ # szl-training-scripts
 
 
35
 
36
+ The forge. Hub copy of the Unsloth-receipted training scripts. Tags were empty — this atelier is the card.
37
 
38
+ **Family.** estate · **Evidence.** HUB · **Weights.** none (UNAVAILABLE on this Hub ID: scripts and receipts only)
39
 
40
+ Hub: [SZLHOLDINGS/szl-training-scripts](https://huggingface.co/SZLHOLDINGS/szl-training-scripts)
41
 
42
+ **Canonical source (GitHub):** the atelier kit lives in [`szl-holdings/a11oy`](https://github.com/szl-holdings/a11oy) under [`.github/atelier-kit/`](https://github.com/szl-holdings/a11oy/tree/main/.github/atelier-kit) (`receipted_unsloth.py`, `unsloth-recipes.json`, `cards/szl-training-scripts.md`, `kit-README.md`). The kit README names [`szl-holdings/szl-khipu`](https://github.com/szl-holdings/szl-khipu), [`szl-holdings/szl-forge`](https://github.com/szl-holdings/szl-forge) and [`szl-holdings/szl-nemo`](https://github.com/szl-holdings/szl-nemo) as its upstream sources. GitHub is canonical; this Hub repository is the mirror.
43
 
44
+ ## Hub tree (OBSERVED 2026-09-25)
 
 
45
 
46
+ Listed through the Hub file API; the revision SHA was not captured by this pass. No weight file of any kind is present.
47
 
48
+ | Path | Bytes | What it is (from the file's own header) |
49
+ |---|---:|---|
50
+ | `receipted_unsloth.py` | 12,165 | "Receipted Unsloth — take the speed, bind the knobs, refuse the folklore." Unsloth QLoRA trainer with one `PROFILES` entry per organ (khipu, khipu-r2, receiptagent, willay, chaski, chaski-5050, chaski-r2), house seed 20260721, `--receipt-only` (mint the recipe receipt, no GPU) and `--list`. Same file name and byte size as the a11oy atelier-kit copy; the first ~3 KB were compared and match, the full-file digest was not compared. |
51
+ | `train_receipted_unsloth.py` | 12,431 | "Receipted Unsloth unique-cut GPU trainer. Runs on Hugging Face Jobs (A10G)." `uv run` script header; `--profile willay` / `--profile chaski`. |
52
+ | `train_willay.py` | 6,799 | WILLAY signed-refusal SFT LoRA script (base `Qwen/Qwen2.5-0.5B-Instruct`, target `SZLHOLDINGS/WILLAY`). |
53
+ | `train_cohort.py` | 17,351 | "CPU-train WILLAY + Chaski silhouettes. Measure kernels. No 1.5B retrain." NumPy silhouettes, seed 20260721, `--kernels-only`. |
54
+ | `szl_khipu_eval.py` | 10,412 | KHIPU-R2 held-out eval port (eval.jsonl 5 navigate + adversarial.jsonl 6; fp16 and optional 4-bit NF4 comparison). |
55
+ | `unsloth-recipes.json` | 16,247 | recipe table (same size as the atelier-kit copy). |
56
+ | `receipts/chaski-unsloth.json`, `receipts/chaski-5050-unsloth.json`, `receipts/chaski-r2-unsloth.json`, `receipts/willay-unsloth.json` | 1,592 / 1,638 / 1,587 / 1,641 | `szl.training_receipt.v2` receipts minted by the profile trainer (unsigned). |
57
+ | `training_receipt.unsloth.json` | 1,638 | profile `chaski-5050` receipt (`path_in_repo: adapter-unsloth`, train loss 0.5588 MEASURED, `evals: none-this-run`, energy UNAVAILABLE, `proven_trust: false`). |
58
+ | `provenance.json` / `status.json` / `LICENSE` | 1,755 / 1,394 / 11,357 | license-candidate records and standalone Apache-2.0 text. `status.json` keeps lineage, consent, privacy, training-suitability and deployment fields BLOCKED and `production_ready: false`. |
59
+ | `bom/model-bom.cdx.json` | 5,324 | CycloneDX BOM |
60
+ | `.gitattributes` | 1,519 | platform-managed |
61
+
62
+ ## The cut
63
+
64
+ Training code is usually a gist. We give it a model id so the estate graph stays one piece.
65
+
66
+ A Hub id you can pin in a receipt: 'trained by this repo at this SHA'.
67
+
68
+ ### Silhouette → leave → SZL
69
+
70
+ | Leader | Take, then tweak |
71
+ |---|---|
72
+ | Anthropic | No public train scripts for Claude. |
73
+ | NVIDIA | NeMo recipes. |
74
+ | Unsloth | The scripts wrap Unsloth v2 profiles. Cut is unique knobs per organ, bound in the receipt. |
75
+
76
+ No equivalent public combination was found among the leaders surveyed above (a snapshot, not an ecosystem-wide novelty claim).
77
+
78
+ ## Intended use
79
+
80
+ Canonical train entrypoint.
81
+
82
+ ## Limitations
83
+
84
+ - Not a checkpoint. Wire to GitHub kit in this atelier.
85
+ - The receipts here are unsigned recipe/run envelopes; a receipt on this Hub ID does not by itself bind any adapter bytes on another repository.
86
+
87
+ ## Honesty
88
+
89
+ | Claim | Label |
90
+ |---|---|
91
+ | This card's numbers | HUB |
92
+ | Energy / joules | UNAVAILABLE unless a signed meter says MEASURED |
93
+ | Λ uniqueness | Conjecture 1 OPEN — not a theorem |
94
+ | GGUF as the signed object | FALSE |
95
+
96
+ Doctrine v11 LOCKED · 749 declarations · 14 axioms · 163 sorries · locked-proven 8.
97
+
98
+ Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr. · ORCID [0009-0001-0110-4173](https://orcid.org/0009-0001-0110-4173).
99
+
100
+ ## GitHub-aligned Python
101
+
102
+ An earlier version of this card embedded a code block that stopped mid-statement (`if trainer.state.log_hi`) and that did not match the `receipted_unsloth.py` actually stored on this Hub ID (it showed an older 2,932-byte single-profile variant with `--base/--r/--seed` flags, which still exists as `atelier/receipted_unsloth.py` in szl-khipu and szl-forge). That block is removed. What follows is quoted verbatim from the 12,165-byte `receipted_unsloth.py` on this repository (2026-09-25 read).
103
+
104
+ Usage lines from the module docstring:
105
+
106
+ ```text
107
+ python receipted_unsloth.py --profile willay --receipt-only
108
+ python receipted_unsloth.py --profile chaski --data doctrine.jsonl --out out/chaski
109
+ ```
110
+
111
+ The receipt-minting tail of `main()` — the point of the file: every knob lands in `training_receipt.json`, loss is read from the trainer log, and the envelope is meant to be signed before any merge:
112
+
113
+ ```python
114
+ trainer = SFTTrainer(model=model, tokenizer=tokenizer, train_dataset=ds, args=SFTConfig(**sft_kw))
115
+ t0 = time.time()
116
+ trainer.train()
117
+ Path(args.out).mkdir(parents=True, exist_ok=True)
118
+ model.save_pretrained(args.out)
119
+ loss = None
120
+ if trainer.state.log_history:
121
+ last = trainer.state.log_history[-1]
122
+ loss = last.get("train_loss") or last.get("loss")
123
+ extra.update({"final_loss": loss, "seconds": round(time.time() - t0, 3), "honesty": "REPORTED"})
124
+ rec = mint_receipt(args.profile, cfg, extra)
125
+ Path(args.out, "training_receipt.json").write_text(json.dumps(rec, indent=2) + "\n")
126
+ print(json.dumps(rec, indent=2))
127
+ return 0
128
+ ```
129
+
130
+ `mint_receipt` refuses `proven_trust: true` and refuses any non-null `energy_j` ("refusing to fabricate joules"); the receipt records `"gguf": "derived — never the signed object"`. Read the full file on this repository or in the canonical kit linked above; do not treat this excerpt as the whole program.
131
 
 
132
 
133
  ## The estate
134
 
 
136
  |---|---|
137
  | Company | [szlholdings.com](https://szlholdings.com) |
138
  | Sponsors | [Become a sponsor](https://szlholdings.com/sponsors.html) |
139
+ | GitHub | [szl-holdings](https://github.com/szl-holdings) |
140
+ | Receipts chain | [frontier-54-runs](https://huggingface.co/datasets/SZLHOLDINGS/frontier-54-runs) |
141
  | Live dashboard | [SI command center](https://szlholdings-szlholdings-si.hf.space) |