VextLabs commited on
Commit
77695ca
·
verified ·
1 Parent(s): 6fa122f

Correct base to juwel-beryl (80L); drop false padding claim; fix load class

Browse files
Files changed (1) hide show
  1. README.md +8 -8
README.md CHANGED
@@ -1,6 +1,6 @@
1
  ---
2
  license: apache-2.0
3
- base_model: VextLabsinc/juwel-sapphire
4
  library_name: peft
5
  pipeline_tag: text-generation
6
  tags:
@@ -12,7 +12,7 @@ tags:
12
  - vext
13
  ---
14
 
15
- > **The adapter weights are hosted in this Hugging Face repo** — load them directly from the Hub. The **base** model (`VextLabsinc/juwel-sapphire`) is currently on Cloudflare R2 (see Download) pending migration to HF: download the base, load it from the local path, then apply this adapter.
16
 
17
  # GEM Ruby · code specialist
18
 
@@ -30,16 +30,16 @@ Never claim first/all weights / AGI / never-forgets / cheapest-ever.
30
 
31
  ## Base model
32
 
33
- - **Required base:** `VextLabsinc/juwel-sapphire`
34
- - **Architecture notes:** Base = JUWEL Sapphire (144-layer base; internal lineage theron-base-v10). This PEFT LoRA is padded from the v9 lineage (80L) to 144L. Not drop-in on stock Qwen2.5-7B.
35
  - **Base license / attribution:** see `NOTICE`
36
 
37
  If this is a **LoRA / PEFT adapter**, load base first, then:
38
 
39
  ```python
40
- from transformers import AutoModelForCausalLM
41
  from peft import PeftModel
42
- base = AutoModelForCausalLM.from_pretrained("VextLabsinc/juwel-sapphire", torch_dtype="auto", device_map="auto")
43
  model = PeftModel.from_pretrained(base, "VextLabsinc/gem-ruby")
44
  ```
45
 
@@ -77,7 +77,7 @@ Vext Labs does **not** endorse misuse. Publishing weights is **not** permission
77
  ## Evaluation
78
 
79
  - **Status:** PENDING — no reproducible benchmark published yet. Evals run on our OWN BF16 weights on a GPU pod (never a hosted API), then posted here with a full audit trail (raw responses, test cases, timestamps, model version). We publish no score we cannot reproduce on our own stack. <!-- e.g. internal rubric 2026-04-02 | pending | public harness -->
80
- - **Numbers:** None yet — honest placeholder, not a hidden result. See docs/ops OPEN_WEIGHTS_BENCHMARK_PLAN.
81
  - Do **not** treat internal rubrics as standardized public leaderboards (MMLU/HELM/etc.) unless re-run under a named public harness.
82
 
83
  ## Files / integrity
@@ -104,4 +104,4 @@ See also `USE_POLICY.md`.
104
  ## Weights — download
105
 
106
  - **Adapter (this model):** hosted in this HF repo — `adapter_model.safetensors` + `adapter_config.json`. Loads with `PeftModel.from_pretrained(base, "VextLabsinc/gem-ruby")`. (R2 mirror: https://pub-a6ae0476e46849f98f1746a61dc4c106.r2.dev/gem-ruby/)
107
- - **Base (`juwel-sapphire`, 144L, BF16):** currently on Cloudflare R2 (migrating to HF). Download all shards from https://pub-a6ae0476e46849f98f1746a61dc4c106.r2.dev/juwel-sapphire/ and load from the local path, then apply this adapter.
 
1
  ---
2
  license: apache-2.0
3
+ base_model: VextLabsinc/juwel-beryl
4
  library_name: peft
5
  pipeline_tag: text-generation
6
  tags:
 
12
  - vext
13
  ---
14
 
15
+ > **Both halves of this model are on the Hub.** The adapter weights live in this repo and the required base is [`VextLabsinc/juwel-beryl`](https://huggingface.co/VextLabsinc/juwel-beryl) (80-layer, BF16). Load the base with `AutoModelForImageTextToText`, then apply this adapter.
16
 
17
  # GEM Ruby · code specialist
18
 
 
30
 
31
  ## Base model
32
 
33
+ - **Required base:** `VextLabsinc/juwel-beryl`
34
+ - **Architecture notes:** Base = JUWEL Beryl (80-layer, hidden 5120, 64 heads, intermediate 25600; internal lineage theron-base-v9 + one CIP rung). This PEFT LoRA is **native to that 80-layer geometry** (`layers_to_transform` 0-79, r=64) — it is not padded and not drop-in on any other JUWEL base. Base architecture is `Qwen3VLForConditionalGeneration` (image-text-to-text), so load it with `AutoModelForImageTextToText`, not `AutoModelForCausalLM`.
35
  - **Base license / attribution:** see `NOTICE`
36
 
37
  If this is a **LoRA / PEFT adapter**, load base first, then:
38
 
39
  ```python
40
+ from transformers import AutoModelForImageTextToText
41
  from peft import PeftModel
42
+ base = AutoModelForImageTextToText.from_pretrained("VextLabsinc/juwel-beryl", torch_dtype="auto", device_map="auto")
43
  model = PeftModel.from_pretrained(base, "VextLabsinc/gem-ruby")
44
  ```
45
 
 
77
  ## Evaluation
78
 
79
  - **Status:** PENDING — no reproducible benchmark published yet. Evals run on our OWN BF16 weights on a GPU pod (never a hosted API), then posted here with a full audit trail (raw responses, test cases, timestamps, model version). We publish no score we cannot reproduce on our own stack. <!-- e.g. internal rubric 2026-04-02 | pending | public harness -->
80
+ - **Numbers:** None yet — honest placeholder, not a hidden result. Benchmarks were blocked until 2026-07-28 by a base-identity error: these cards named a 144-layer base for an 80-layer adapter, so the adapter was not loadable as trained. The base identity is now corrected and evaluation is under way.
81
  - Do **not** treat internal rubrics as standardized public leaderboards (MMLU/HELM/etc.) unless re-run under a named public harness.
82
 
83
  ## Files / integrity
 
104
  ## Weights — download
105
 
106
  - **Adapter (this model):** hosted in this HF repo — `adapter_model.safetensors` + `adapter_config.json`. Loads with `PeftModel.from_pretrained(base, "VextLabsinc/gem-ruby")`. (R2 mirror: https://pub-a6ae0476e46849f98f1746a61dc4c106.r2.dev/gem-ruby/)
107
+ - **Base (`juwel-beryl`, 80L, BF16):** on the Hub at https://huggingface.co/VextLabsinc/juwel-beryl — loads directly by repo id, no manual download needed.