docs(card): honest ROADMAP silhouette — no weights, no fake base_model

#2
by betterwithage - opened
Files changed (1) hide show
  1. README.md +28 -45
README.md CHANGED
@@ -1,67 +1,50 @@
1
  ---
2
  license: apache-2.0
3
- language:
4
- - en
5
- pipeline_tag: feature-extraction
6
- library_name: sentence-transformers
7
- base_model: Qwen/Qwen3-Embedding-0.6B
8
- base_model_relation: finetune
9
  tags:
 
 
10
  - szl-holdings
11
  - series-a
12
  - doctrine-v11
13
- - embeddings
14
- - retrieval
15
- - roadmap
16
- szl:
17
- doctrine: v11-LOCKED
18
- lean: "749/14/163"
19
- lambda: "Conjecture 1 — advisory, never a theorem"
20
- evidence_ceiling: 0.97
21
- artifact_class: WEIGHTS
22
- originality: FINETUNE_DISCLOSED_BASE
23
- collection: "SZL Fall 2026 — Original Cuts"
24
- jobs: UNKNOWN
25
  ---
26
- ROADMAP.
27
 
28
  # Chakana
29
 
30
- **One line.** Bridge embeddings for szl-lake / doctrine / killinchu text retrieval.
 
 
 
 
31
 
32
- | | |
33
- |---|---|
34
- | **Artifact** | encoder (none uploaded yet) |
35
- | **Originality** | sentence-transformers fine-tune of a disclosed Apache embedding base |
36
- | **Base** | `Qwen/Qwen3-Embedding-0.6B` (Apache-2.0). Alternate MIT: `BAAI/bge-m3`. Never KaLM. Never EmbeddingGemma. |
37
- | **License** | `apache-2.0` |
38
- | **Demo** | ROADMAP. MiniEmbed inside `szl-kernels` is a different, smaller co-occurrence artefact. |
39
- | **Status** | CUTTING · ROADMAP |
40
- | **HF Jobs** | UNKNOWN (no job id) |
41
- | **Sibling** | [`tinku`](https://huggingface.co/SZLHOLDINGS/tinku) (reranker, same freeze split) |
42
 
43
- > **Fashion rule.** Silhouette from Qwen3-Embedding / MTEB leaders. Cut is original SZL pairs. We do not republish Qwen3-Embedding tensors as ours. We do not paste their MTEB numbers.
44
 
45
- ## Intended use
 
 
46
 
47
- - **Who:** a11oy retrieval over the proof registry
48
- - **What:** query → vector; cosine retrieve; optional Matryoshka 256/512/1024
49
- - **Where:** behind a controller. This is not an autonomous searcher.
50
 
51
- ## What it is NOT
 
 
 
 
52
 
53
- - Not MiniEmbed renamed. Not a Qwen3-Embedding rehost. Not KaLM. Not a leaderboard clone.
54
 
55
- ## Evaluation
56
 
57
- **Status: UNKNOWN.** Protocol when a job runs: MTEB subset we actually execute, plus a frozen in-house nDCG@10. If we have not run MTEB, this card stays UNKNOWN. No 64.34 / 70.58 / 72.32 paste.
 
 
58
 
59
- ## Training
60
 
61
- - **Data:** only rights-admitted SZL pairs (`szl-lake`, `rag-corpus-v1`) if QILLQA/FORGE confirm admission. Else SKIP.
62
- - **Train receipt:** absent.
63
- - **Jobs:** UNKNOWN until a run id exists. Owner-metal Unsloth so far is other models, REPORTED, not this one.
64
 
65
- ## Limitations
 
66
 
67
- - No weights yet. Λ = Conjecture 1. Trust ceiling 0.97.
 
1
  ---
2
  license: apache-2.0
 
 
 
 
 
 
3
  tags:
4
+ - roadmap
5
+ - not-a-checkpoint
6
  - szl-holdings
7
  - series-a
8
  - doctrine-v11
 
 
 
 
 
 
 
 
 
 
 
 
9
  ---
 
10
 
11
  # Chakana
12
 
13
+ **ROADMAP.** This Hub ID is a named Series-A slot, not a trained model. There are **no weights**. Do not `from_pretrained` this ID. Nothing here will load.
14
+
15
+ Chakana is the named Series-A embeddings slot. **ROADMAP**, no checkpoint.
16
+
17
+ ## What
18
 
19
+ - A reserved name for a future embeddings checkpoint.
20
+ - Product intent: vectors for house retrieval. No encoder lives on this ID.
 
 
 
 
 
 
 
 
21
 
22
+ ## What-NOT
23
 
24
+ - **Not a Qwen3-Embedding fine-tune.** Not `base_model: Qwen/Qwen3-Embedding-0.6B`.
25
+ - Not MiniEmbed renamed.
26
+ - Not an MTEB clone. We do not paste anyone else's retrieval numbers.
27
 
28
+ ## Status
 
 
29
 
30
+ | Thing | Label | Method / N / date / what-NOT |
31
+ |---|---|---|
32
+ | Weights | **UNAVAILABLE** | Hub file list 2026-08-28 1:18 PM ET. N=0 weight files (README + `.gitattributes` only). What-NOT: this is not a checkpoint. |
33
+ | Evals | **ROADMAP** | No eval receipt on this ID. N=0. What-NOT: no accuracy, no latency, no leaderboard paste. |
34
+ | Inference | **UNAVAILABLE** | `from_pretrained` this ID will fail. |
35
 
36
+ ## Live weights (different IDs)
37
 
38
+ Tensors that actually load today:
39
 
40
+ - [`SZLHOLDINGS/SZL-Khipu-1.5B`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B) and [`SZL-Khipu-1.5B-GGUF`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B-GGUF)
41
+ - [`SZLHOLDINGS/SZL-Forge-1.5B-ReceiptAgent`](https://huggingface.co/SZLHOLDINGS/SZL-Forge-1.5B-ReceiptAgent)
42
+ - [`SZLHOLDINGS/chaski`](https://huggingface.co/SZLHOLDINGS/chaski) — adapter + merged 16-bit on Hub; evals none-this-run; not production
43
 
44
+ ## Doctrine
45
 
46
+ **Doctrine v11.** Λ (lambda) = **Conjecture 1** advisory, never green, never a theorem.
 
 
47
 
48
+ Honesty labels: **MEASURED** · **REPORTED** · **DERIVED** · **MODELED** · **SNAPSHOT** · **REACHABLE** · **UNAVAILABLE** · **ROADMAP** · **CONJECTURE**.
49
+ Every number carries method / N / date / what-NOT. No model-index. No invented evals.
50