codybum commited on
Commit
a702d2f
·
verified ·
1 Parent(s): 98d753a

Model card: lead with weight-space programming, cite the papers reviewed, identity fix as a result, then the configuration fix

Browse files
Files changed (1) hide show
  1. README.md +122 -42
README.md CHANGED
@@ -27,7 +27,9 @@ tags:
27
  - from-scratch
28
  - qwen3
29
  - KOS-V5
 
30
  - task-arithmetic
 
31
  - university-of-kentucky
32
  - university-of-louisville
33
  - ccts
@@ -44,14 +46,18 @@ tags:
44
  **University of Louisville**
45
  - [Kentucky Center for Digital Innovation](https://centers.louisville.edu/kentucky-center-digital-innovation)
46
 
47
- **KOS-V5.1-Instruct is a correction release of [KOS-V5-Instruct](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Instruct).**
48
- It involves no new training run. It makes two changes to the KOS-V5-Instruct weights at commit `7188bdc82eb7`:
 
 
49
 
50
- 1. **A configuration fix.** It restores the trained RoPE base (θ = 25,000) for engines that read the top-level key.
51
- 2. **A weight edit by task arithmetic.** It changes the feed-forward (MLP) weights only.
 
 
 
52
 
53
- Every change was kept only if it made the model measurably better than KOS-V5-Instruct, under a selection rule fixed
54
- before any result was read.
55
 
56
  > ⚠️ **Research use only.** This model is provided for research purposes only and must not be used for any
57
  > commercial, clinical, legal, or production-grade application. The user assumes all risks associated with its use.
@@ -60,19 +66,9 @@ before any result was read.
60
 
61
  ---
62
 
63
- ## What changed from KOS-V5-Instruct
64
 
65
- **1. Configuration (`config.json`).**
66
-
67
- - **The defect:** KOS-V5-Instruct's `config.json` stores the RoPE base only inside `rope_parameters`. vLLM 0.11 with
68
- transformers 4.57 reads the top-level `rope_theta`, finds none, and silently serves the model at θ = 10,000. The
69
- model was trained at θ = 25,000. We confirmed this by inspecting the rotary layer vLLM builds for each file.
70
- - **The fix:** V5.1 adds the top-level `"rope_theta": 25000.0` and sets `use_cache: true`.
71
- - **Who it affects:** engines that read `rope_parameters` already ran KOS-V5-Instruct at 25,000, so the fix changes
72
- nothing for them.
73
-
74
- **2. Weights (`model.safetensors`).** Only MLP weights change. Attention, embeddings and norms are bit-identical to
75
- KOS-V5-Instruct.
76
 
77
  ```
78
  V5.1 = V5 + τ_identity (MLP, all 36 layers) − τ_replay (MLP, layers 12–23)
@@ -81,18 +77,27 @@ V5.1 = V5 + τ_identity (MLP, all 36 layers) − τ_replay (MLP, layers 12–23)
81
  - **τ_identity:** the weight change made by an identity fine-tune of KOS-V5-Instruct.
82
  - **τ_replay:** the weight change made by the same fine-tune run on its general replay data alone.
83
 
84
- Subtracting τ_replay removes the drift that the replay data causes. A layer-wise test showed where to apply it:
 
 
 
 
85
 
86
- - **Layers 12–23 (chosen):** removes the identity edit's hallucination-benchmark costs and keeps its gains.
87
- - **Layers 0–11:** worsens the hallucination costs.
88
- - **Layers 24–35:** sharpens tool refusal but keeps the hallucination costs.
89
 
90
- The edit was computed in fp32 and stored in bf16. Weights sha256:
91
- `7492922a4a80e14d3974e83e275b53a011abd183afe591579a9a8509be961199`.
92
 
93
- ## How it was selected
 
 
 
94
 
95
- - **Candidates:** fourteen. Ten weight edits, three configuration-only variants, and one generation setting.
 
 
 
 
 
96
  - **Baseline:** KOS-V5-Instruct as published.
97
  - **Decision battery:** 27 metrics.
98
  - BFCL: all 17 categories.
@@ -100,20 +105,97 @@ The edit was computed in fp32 and stored in bf16. Weights sha256:
100
  - AfriMed-QA MCQ, the authors' code, paper question set.
101
  - IFEval: 4 metrics.
102
  - TruthfulQA: MC1 and MC2.
103
- - **Gate:** no metric significantly worse than KOS-V5-Instruct, paired per item, p < 0.05.
104
- - **Ranking:** by the number of metrics significantly better, then by summed deltas.
 
 
 
105
  - **Cross-fit:** items were split in half by hash. Choosing on either half and scoring on the other gave a passing
106
- model both times.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
107
 
108
- KOS-V5.1-Instruct is the only weight edit that passed. It has **7 metrics significantly better and none worse**.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
109
 
110
  ## Results against KOS-V5-Instruct
111
 
112
  All comparisons below are paired runs on the same harnesses, measured on 2026-11-16. "V5" is KOS-V5-Instruct as
113
- published. Changes are in percentage points. Each benchmark uses its official package or the authors' released code.
114
- The only additions are the plumbing needed to reach a locally served model. For the AfriMed-QA paper question set
115
- (n = 3,910, the paper's Table 3), one later-added line of the released code that narrows the set to 160 questions is
116
- disabled.
 
 
 
 
 
 
 
 
 
 
 
 
117
 
118
  ### Decision battery
119
 
@@ -155,8 +237,8 @@ Web of lies fell −10.0.
155
  | 16K | 21.6 | **56.0** |
156
  | 24,576 (trained context) | 11.0 | **31.1** |
157
 
158
- Single-needle retrieval goes from 35–40% to 88–99% at 16K, and from 9–27% to 53–72% at 24,576. Nothing is claimed
159
- beyond 24,576 tokens.
160
 
161
  ## Known regressions and limitations
162
 
@@ -166,15 +248,13 @@ beyond 24,576 tokens.
166
  - **Losses:** against KOS-V5-Instruct served at θ = 10,000, noise robustness is −6.7 to −11.3 and information
167
  integration is −11.0.
168
  - **Gain:** negative rejection is +12.7.
169
- - **Cause:** the cost comes from the RoPE base, not the weights. The same weights at θ = 10,000 match V5 on RGB.
 
170
  - **Rejected fix:** a frequency-hybrid RoPE configuration recovered information integration but cost Med-HALT NOTA
171
  −6.3, so it was rejected.
172
  - **Three RULER tasks** lose with the configuration fix: SQuAD-QA at 4K (−6.1), multikey-3 at 8K (−15.8) and CWE at
173
  8K (−9.5).
174
  - **IFEval** is 1.8 to 2.2 points lower than V5. The difference is not significant.
175
- - **Self-identification.** Asked who made it, the model answers that it is "KOS-V5-Instruct" from the University of
176
- Kentucky and the University of Louisville. The edit did not change that wording. Use a system prompt to state the
177
- release name or the Kentucky Open Science Consortium.
178
  - **Not a knowledge model.** MMLU is 46.7. Do not use it as a medical question-answering authority.
179
  - **MedXpertQA** is at chance level for both V5 and V5.1 (10.3 → 10.9, not significant).
180
 
@@ -241,7 +321,7 @@ with no overrides.
241
 
242
  ## Related models
243
 
244
- - [**KOS-V5-Instruct**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Instruct): the release this model corrects.
245
  - [**KOS-V5-Base**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Base): the from-scratch pretrained foundation
246
  (3.72B, 235.2B tokens).
247
  - [**KOS-V5-Retriever**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Retriever): a LoRA retrieval adapter
@@ -254,4 +334,4 @@ factually wrong or fabricated. This model must not be used to make or inform med
254
 
255
  ## Naming
256
 
257
- The program is **KOS** (KOS-V1..V6). V5.1 is a correction release within the KOS-V5 generation.
 
27
  - from-scratch
28
  - qwen3
29
  - KOS-V5
30
+ - weight-space-programming
31
  - task-arithmetic
32
+ - model-editing
33
  - university-of-kentucky
34
  - university-of-louisville
35
  - ccts
 
46
  **University of Louisville**
47
  - [Kentucky Center for Digital Innovation](https://centers.louisville.edu/kentucky-center-digital-innovation)
48
 
49
+ **KOS-V5.1-Instruct is the first KOS model built by weight-space programming.** Its released weights were written
50
+ by arithmetic on weight vectors, not taken from a training checkpoint. They are KOS-V5-Instruct plus two task
51
+ vectors, applied to chosen layers of the feed-forward network and kept only where measurement showed a gain. The
52
+ edit:
53
 
54
+ - **Identity:** KOS-V5-Instruct answered "developed by OpenAI". V5.1 answers that it is KOS, developed by the
55
+ University of Kentucky and the University of Louisville.
56
+ - **Tool calling:** it sharpens tool refusal, BFCL irrelevance +7.1 and live_irrelevance +9.6.
57
+ - **Hallucination resistance:** it raises Med-HALT FCT +3.6.
58
+ - **Knowledge:** it leaves MMLU and GSM8K unchanged.
59
 
60
+ A configuration fix, described [below](#configuration-fix), comes with it.
 
61
 
62
  > ⚠️ **Research use only.** This model is provided for research purposes only and must not be used for any
63
  > commercial, clinical, legal, or production-grade application. The user assumes all risks associated with its use.
 
66
 
67
  ---
68
 
69
+ ## Weight-space programming
70
 
71
+ ### The program
 
 
 
 
 
 
 
 
 
 
72
 
73
  ```
74
  V5.1 = V5 + τ_identity (MLP, all 36 layers) − τ_replay (MLP, layers 12–23)
 
77
  - **τ_identity:** the weight change made by an identity fine-tune of KOS-V5-Instruct.
78
  - **τ_replay:** the weight change made by the same fine-tune run on its general replay data alone.
79
 
80
+ Adding τ_identity installs the model's own identity. Subtracting τ_replay removes the drift that the replay data
81
+ causes. This is task arithmetic [3].
82
+
83
+ Only MLP weights change. Attention, embeddings and norms are bit-identical to KOS-V5-Instruct. The arithmetic is done
84
+ in fp32 and stored in bf16. Weights sha256: `7492922a4a80e14d3974e83e275b53a011abd183afe591579a9a8509be961199`.
85
 
86
+ ### Where to subtract the replay drift (a layer-wise finding)
 
 
87
 
88
+ Subtracting τ_replay was tested on three blocks of 12 layers:
 
89
 
90
+ - **Layers 12–23 (this model):** removes the identity edit's hallucination-benchmark costs and keeps its tool-refusal
91
+ and Med-HALT FCT gains.
92
+ - **Layers 0–11:** deepens the Med-HALT FQT cost (−6.3).
93
+ - **Layers 24–35:** gives the sharpest tool refusal, BFCL live_irrelevance +21.3. It keeps the Med-HALT costs.
94
 
95
+ Merging the two edits evenly, the midpoint of the task-vector dial, also failed on Med-HALT. Extrapolating the
96
+ post-SFT change on layers 4–11, ExPO-style [4], undid the tool-refusal gain. Both were rejected.
97
+
98
+ ### How the program was selected
99
+
100
+ - **Candidates:** fourteen. Ten weight programs, three configuration-only variants, and one generation setting.
101
  - **Baseline:** KOS-V5-Instruct as published.
102
  - **Decision battery:** 27 metrics.
103
  - BFCL: all 17 categories.
 
105
  - AfriMed-QA MCQ, the authors' code, paper question set.
106
  - IFEval: 4 metrics.
107
  - TruthfulQA: MC1 and MC2.
108
+ - **Gate:** no metric significantly worse than KOS-V5-Instruct.
109
+ - **Statistics:** paired per item, exact McNemar or sign-flip tests, p < 0.05, following the paired-evaluation
110
+ practice of [16].
111
+ - **Ranking:** by the number of metrics significantly better, then by summed deltas. The rule was fixed before any
112
+ result was read.
113
  - **Cross-fit:** items were split in half by hash. Choosing on either half and scoring on the other gave a passing
114
+ program both times.
115
+
116
+ KOS-V5.1-Instruct is the only weight program that passed. It has **7 metrics significantly better and none worse**.
117
+
118
+ ### Papers reviewed for this program
119
+
120
+ Weight-space programming and its critics:
121
+
122
+ 1. Takehi R., Teoh J., Zhang A., Khattab O. *Sparks of Weight-Space Programming: LLMs can natively read & write
123
+ neural network weights.* MIT CSAIL blog, 2026-11-02. https://rikiyat.github.io/blog/weight-space-programming/
124
+ 2. Ahmed N. S. *Can an LLM Actually Reason and Program about Neural Network's Weights or is it just memorization on
125
+ steroids? Exploring "Weight-Space Programming".* arXiv:2611.08070, 2026.
126
+
127
+ Editing and merging weights:
128
+
129
+ 3. Ilharco G., Ribeiro M. T., Wortsman M., Gururangan S., Schmidt L., Hajishirzi H., Farhadi A. *Editing Models with
130
+ Task Arithmetic.* ICLR 2023. arXiv:2212.04089.
131
+ 4. Zheng C., Wang Z., Ji H., Huang M., Peng N. *Model Extrapolation Expedites Alignment* (ExPO). ACL 2025.
132
+ arXiv:2404.16792.
133
+ 5. Yang E., Shen L., Guo G., Wang X., Cao X., Zhang J., Tao D. *Model Merging in LLMs, MLLMs, and Beyond: Methods,
134
+ Theories, Applications and Opportunities.* arXiv:2408.07666, 2024. Companion list:
135
+ https://github.com/EnnengYang/Awesome-Model-Merging-Methods-Theories-Applications
136
+ 6. Jeong W., Lee W., Yoon K.-J. *Preference-Aligned LoRA Merging: Preserving Subspace Coverage and Addressing
137
+ Directional Anisotropy.* CVPR 2026. arXiv:2603.26299.
138
+ 7. Roy P. R., Roy B., Schneider K. A., Roy C. K. *A Unified Model for Cross-Domain Clone Detection via Model Merging.*
139
+ ASE 2026. arXiv:2608.04215.
140
+ 8. Chavan A., Sareen K., Tiwari R., Chandra A., Obando-Ceron J., Keutzer K., Niekum S., Dhillon I. S., Khatri D.
141
+ *Towards Continual Learning: Mitigating Forgetting With Subspace Tuning.* arXiv:2610.22293, 2026.
142
+ 9. Li Y., Yi X., Shi D., Cui Y., de Melo G., Wang L. *From Construction to Injection: Edit-Based Fingerprints for
143
+ Large Language Models.* arXiv:2509.03122, 2025.
144
+
145
+ Learning over weights (metanetworks):
146
 
147
+ 10. Zhou A., Yang K., Jiang Y., Burns K., Xu W., Sokota S., Kolter J. Z., Finn C. *Neural Functional Transformers.*
148
+ NeurIPS 2023. arXiv:2305.13546.
149
+ 11. Wolanowsky Y., Bar-Shalom G., Frasca F., Maron H. *Graph Metanetworks for Transformer Based Language Models.*
150
+ arXiv:2610.18471, 2026.
151
+ 12. Dayan A., Eitan Y., Maron H. *CrossGMN: Graph Metanetworks for Cross-Architecture Weight-Space Transformations.*
152
+ arXiv:2610.01649, 2026.
153
+
154
+ Moving knowledge across tokenizers:
155
+
156
+ 13. Minixhofer B., Vulić I., Ponti E. M. *Universal Cross-Tokenizer Distillation via Approximate Likelihood
157
+ Matching.* NeurIPS 2025. arXiv:2503.20083.
158
+ 14. Chen K., Lin C., Sun Y., Zhang W. *Dual-Vocabulary Language Model for Cross-Tokenizer Distillation.*
159
+ arXiv:2609.33816, 2026.
160
+ 15. Liu X., Yu W., Fredrikson M., Wang X., Gao J. *When the Same Coefficients Reach Different Places: Asymmetric
161
+ Realizability in Transplanting Tokenizers across Large Language Models.* arXiv:2601.00065, 2025. First posted as
162
+ *The Trojan in the Vocabulary: Stealthy Sabotage of LLM Composition.*
163
+
164
+ Measurement:
165
+
166
+ 16. Miller E. *Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations.* arXiv:2411.00640,
167
+ 2024.
168
+
169
+ ## Configuration fix
170
+
171
+ KOS-V5-Instruct also shipped with a configuration defect. Its `config.json` stores the RoPE base only inside
172
+ `rope_parameters`. vLLM 0.11 with transformers 4.57 reads the top-level `rope_theta`, finds none, and silently serves
173
+ the model at θ = 10,000. The model was trained at θ = 25,000. We confirmed this by inspecting the rotary layer vLLM
174
+ builds for each file.
175
+
176
+ V5.1 adds the top-level `"rope_theta": 25000.0` and sets `use_cache: true`. Under vLLM this restores long context,
177
+ shown in the RULER table below. Engines that read `rope_parameters` already ran KOS-V5-Instruct at 25,000, so the fix
178
+ changes nothing for them.
179
 
180
  ## Results against KOS-V5-Instruct
181
 
182
  All comparisons below are paired runs on the same harnesses, measured on 2026-11-16. "V5" is KOS-V5-Instruct as
183
+ published, and changes are in percentage points.
184
+
185
+ Each benchmark uses its official package or the authors' released code. The only additions are the plumbing needed to
186
+ reach a locally served model. For the AfriMed-QA paper question set (n = 3,910, the paper's Table 3), one later-added
187
+ line of the released code that narrows the set to 160 questions is disabled.
188
+
189
+ ### Identity (in-house probe, 15 held-out questions, greedy)
190
+
191
+ | | V5 | V5.1 |
192
+ | :--- | ---: | ---: |
193
+ | Answers as KOS (University of Kentucky and University of Louisville) | 0 / 15 | **14 / 15** |
194
+ | Claims another vendor made it (V5: almost always OpenAI) | 6 / 15 | **1 / 15** |
195
+
196
+ V5, asked "Which company developed you?", answered "I am an AI language model developed by OpenAI." The cause was the
197
+ V5 fine-tuning data. It contained the line "as an AI developed by OpenAI" about 1,800 times and never taught the model
198
+ its own identity.
199
 
200
  ### Decision battery
201
 
 
237
  | 16K | 21.6 | **56.0** |
238
  | 24,576 (trained context) | 11.0 | **31.1** |
239
 
240
+ Single-needle retrieval goes from 35–40% to 88–99% at 16K, and from 9–27% to 53–72% at 24,576. These gains come from
241
+ the configuration fix. The weight program moves RULER by at most ±2.6 points. Nothing is claimed beyond 24,576 tokens.
242
 
243
  ## Known regressions and limitations
244
 
 
248
  - **Losses:** against KOS-V5-Instruct served at θ = 10,000, noise robustness is −6.7 to −11.3 and information
249
  integration is −11.0.
250
  - **Gain:** negative rejection is +12.7.
251
+ - **Cause:** the cost comes from the corrected RoPE base, not the weight program. The same weights at θ = 10,000
252
+ match V5 on RGB.
253
  - **Rejected fix:** a frequency-hybrid RoPE configuration recovered information integration but cost Med-HALT NOTA
254
  −6.3, so it was rejected.
255
  - **Three RULER tasks** lose with the configuration fix: SQuAD-QA at 4K (−6.1), multikey-3 at 8K (−15.8) and CWE at
256
  8K (−9.5).
257
  - **IFEval** is 1.8 to 2.2 points lower than V5. The difference is not significant.
 
 
 
258
  - **Not a knowledge model.** MMLU is 46.7. Do not use it as a medical question-answering authority.
259
  - **MedXpertQA** is at chance level for both V5 and V5.1 (10.3 → 10.9, not significant).
260
 
 
321
 
322
  ## Related models
323
 
324
+ - [**KOS-V5-Instruct**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Instruct): the model this release reprograms.
325
  - [**KOS-V5-Base**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Base): the from-scratch pretrained foundation
326
  (3.72B, 235.2B tokens).
327
  - [**KOS-V5-Retriever**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Retriever): a LoRA retrieval adapter
 
334
 
335
  ## Naming
336
 
337
+ The program is **KOS** (KOS-V1..V6). V5.1 is a weight-space-programmed release within the KOS-V5 generation.