Humuhumu33 commited on
Commit
6504dcd
verified
1 Parent(s): 08f65db

Add branded Hologram model card

Browse files
Files changed (1) hide show
  1. README.md +83 -0
README.md ADDED
@@ -0,0 +1,83 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ base_model: Dream-org/Dream-v0-Instruct-7B
4
+ pipeline_tag: text-generation
5
+ inference: false
6
+ tags:
7
+ - hologram
8
+ - q
9
+ - kappa-object
10
+ - holo
11
+ - diffusion
12
+ - qwen2
13
+ ---
14
+
15
+ <div align="center">
16
+
17
+ # Hologram 路 Dream-7B
18
+
19
+ **Diffusion language model**
20
+
21
+ `q3f 路 diffusion` 路 `2.9 GB` 路 streamed to Q as a **key-addressable `.holo` object**
22
+
23
+ [Hologram](https://gethologram.ai) 路 [Live Space](https://huggingface.co/spaces/HOLOGRAMTECH/hologram) 路 [Organization](https://huggingface.co/HOLOGRAMTECH) 路 [Code](https://github.com/Hologram-Technologies)
24
+
25
+ </div>
26
+
27
+ ---
28
+
29
+ ## What this is
30
+
31
+ A masked-diffusion LM: generation is iterative bidirectional unmasking over denoising steps, not left to right. Wall-clock is fixed by step count, not output length.
32
+
33
+ This repository is **not** a GGUF or Transformers checkpoint. It is a **Hologram key object**: the weights of `Dream-org/Dream-v0-Instruct-7B` re-encoded into Hologram's content-addressed `.holo` format so they stream, one verified block at a time, into **Q**, the on-device brain of the Hologram web OS. It runs in the browser on WebGPU, serverless, with nothing to install.
34
+
35
+ ## How it streams
36
+
37
+ The object is laid out for cold streaming from an untrusted CDN:
38
+
39
+ | File | Role |
40
+ |---|---|
41
+ | `manifest.json` | the root. Names every tensor and the key (content hash) of its block. |
42
+ | `b/sha256_*.gz` | the tensor blocks. Each filename **is** the SHA-256 of its bytes. |
43
+ | `tokenizer.gguf` | bundled header (where present), so loading is fully serverless. |
44
+
45
+ Q fetches the manifest, then pulls each block by its key and re-derives `sha256(block)` on arrival. If a byte is wrong, the block is rejected. Nothing is trusted; everything is proven.
46
+
47
+ ## Verify (Law L5)
48
+
49
+ The object's identity is the SHA-256 of its manifest, pinned in Q's catalog before a single byte of weight is trusted:
50
+
51
+ ```
52
+ did:holo:sha256:7b862931ae088f348f1f7e9ea3adbd418924c2e07e6ddd134f926e5681ad760d
53
+ ```
54
+
55
+ ```bash
56
+ # the manifest hash equals the pinned identity above
57
+ curl -sL https://huggingface.co/HOLOGRAMTECH/q-dream-7b/resolve/main/manifest.json | sha256sum
58
+ ```
59
+
60
+ ## Specifications
61
+
62
+ | | |
63
+ |---|---|
64
+ | Architecture | Qwen2.5-7B backbone (masked diffusion) |
65
+ | Precision | q3f 路 diffusion |
66
+ | Object size | 2.9 GB |
67
+ | Hidden size | 3584 |
68
+ | Layers | 28 |
69
+ | Heads (Q / KV) | 28 / 4 (GQA) |
70
+ | FFN | 18944 |
71
+ | Vocab | 152064 |
72
+ | Context | 192 |
73
+ | Format | `holo-2bit/1` |
74
+
75
+ ## Provenance and license
76
+
77
+ Derived from [`Dream-org/Dream-v0-Instruct-7B`](https://huggingface.co/Dream-org/Dream-v0-Instruct-7B). Inherits its license from Dream-org/Dream-v0-Instruct-7B. The re-encoding is lossless-by-construction at the key level: every block is content-addressed, so the object either re-derives to its pinned identity or it is refused.
78
+
79
+ ## Run it
80
+
81
+ These weights load through Q, not a standard runtime. Open the [Live Space](https://huggingface.co/spaces/HOLOGRAMTECH/hologram) or visit [gethologram.ai](https://gethologram.ai) to run Hologram, then pick **Dream-7B** from Q's model list.
82
+
83
+ <div align="center"><sub>Composed on the golden ratio. One key, everything.</sub></div>