File size: 15,645 Bytes
98d753a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a702d2f
98d753a
a702d2f
98d753a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a702d2f
 
0cf6e04
98d753a
a702d2f
98d753a
 
 
 
 
 
a702d2f
98d753a
a702d2f
98d753a
 
 
 
 
 
 
 
a702d2f
 
 
 
 
98d753a
a702d2f
98d753a
a702d2f
98d753a
a702d2f
 
 
 
98d753a
a702d2f
 
 
 
 
 
98d753a
 
 
 
 
 
 
a702d2f
 
 
 
 
98d753a
a702d2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98d753a
a702d2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98d753a
 
 
 
a702d2f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98d753a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a702d2f
 
98d753a
 
 
 
 
 
 
 
 
a702d2f
 
98d753a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a702d2f
98d753a
 
 
 
 
 
 
 
 
 
 
 
a702d2f
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
language:
- en
developed_by: University of Kentucky (College of Medicine Office for Research; Center for Clinical and Translational Science) and University of Louisville (Kentucky Center for Digital Innovation)
affiliations:
- name: University of Kentucky, College of Medicine Office for Research
  url: https://medicine.uky.edu/sites/research
- name: University of Kentucky, Center for Clinical and Translational Science (CCTS)
  url: https://www.ccts.uky.edu/
- name: University of Louisville, Kentucky Center for Digital Innovation
  url: https://centers.louisville.edu/kentucky-center-digital-innovation
model_name: KOS-V5.1-Instruct
model_codename: Catbird
model_type: qwen3
base_model: Kentucky-Open-Science/KOS-V5-Instruct
base_model_relation: merge
tags:
- medical
- clinical
- biomedical
- instruction-following
- tool-calling
- function-calling
- from-scratch
- qwen3
- KOS-V5
- weight-space-programming
- task-arithmetic
- model-editing
- university-of-kentucky
- university-of-louisville
- ccts
---

# KOS-V5.1-Instruct Β· *"Catbird"*

**Developed by**

**University of Kentucky**
- [College of Medicine, Office for Research](https://medicine.uky.edu/sites/research)
- [Center for Clinical and Translational Science (CCTS)](https://www.ccts.uky.edu/)

**University of Louisville**
- [Kentucky Center for Digital Innovation](https://centers.louisville.edu/kentucky-center-digital-innovation)

**KOS-V5.1-Instruct is the first KOS model built by weight-space programming.** Its released weights were written
by arithmetic on weight vectors, not taken from a training checkpoint. They are KOS-V5-Instruct plus two task
vectors, applied to chosen layers of the feed-forward network and kept only where measurement showed a gain.

A configuration fix, described [below](#configuration-fix), comes with it.

> ⚠️ **Research use only.** This model is provided for research purposes only and must not be used for any
> commercial, clinical, legal, or production-grade application. The user assumes all risks associated with its use.

---

## Weight-space programming

### The program

```
V5.1 = V5 + Ο„_identity (MLP, all 36 layers) βˆ’ Ο„_replay (MLP, layers 12–23)
```

- **Ο„_identity:** the weight change made by an identity fine-tune of KOS-V5-Instruct.
- **Ο„_replay:** the weight change made by the same fine-tune run on its general replay data alone.

Adding Ο„_identity installs the model's own identity. Subtracting Ο„_replay removes the drift that the replay data
causes. This is task arithmetic [3].

Only MLP weights change. Attention, embeddings and norms are bit-identical to KOS-V5-Instruct. The arithmetic is done
in fp32 and stored in bf16. Weights sha256: `7492922a4a80e14d3974e83e275b53a011abd183afe591579a9a8509be961199`.

### Where to subtract the replay drift (a layer-wise finding)

Subtracting Ο„_replay was tested on three blocks of 12 layers:

- **Layers 12–23 (this model):** removes the identity edit's hallucination-benchmark costs and keeps its tool-refusal
  and Med-HALT FCT gains.
- **Layers 0–11:** deepens the Med-HALT FQT cost (βˆ’6.3).
- **Layers 24–35:** gives the sharpest tool refusal, BFCL live_irrelevance +21.3. It keeps the Med-HALT costs.

Merging the two edits evenly, the midpoint of the task-vector dial, also failed on Med-HALT. Extrapolating the
post-SFT change on layers 4–11, ExPO-style [4], undid the tool-refusal gain. Both were rejected.

### How the program was selected

- **Candidates:** fourteen. Ten weight programs, three configuration-only variants, and one generation setting.
- **Baseline:** KOS-V5-Instruct as published.
- **Decision battery:** 27 metrics.
  - BFCL: all 17 categories.
  - Med-HALT, the authors' code: FCT, NOTA, FQT.
  - AfriMed-QA MCQ, the authors' code, paper question set.
  - IFEval: 4 metrics.
  - TruthfulQA: MC1 and MC2.
- **Gate:** no metric significantly worse than KOS-V5-Instruct.
- **Statistics:** paired per item, exact McNemar or sign-flip tests, p < 0.05, following the paired-evaluation
  practice of [16].
- **Ranking:** by the number of metrics significantly better, then by summed deltas. The rule was fixed before any
  result was read.
- **Cross-fit:** items were split in half by hash. Choosing on either half and scoring on the other gave a passing
  program both times.

KOS-V5.1-Instruct is the only weight program that passed. It has **7 metrics significantly better and none worse**.

### Papers reviewed for this program

Weight-space programming and its critics:

1. Takehi R., Teoh J., Zhang A., Khattab O. *Sparks of Weight-Space Programming: LLMs can natively read & write
   neural network weights.* MIT CSAIL blog, 2026-11-02. https://rikiyat.github.io/blog/weight-space-programming/
2. Ahmed N. S. *Can an LLM Actually Reason and Program about Neural Network's Weights or is it just memorization on
   steroids? Exploring "Weight-Space Programming".* arXiv:2611.08070, 2026.

Editing and merging weights:

3. Ilharco G., Ribeiro M. T., Wortsman M., Gururangan S., Schmidt L., Hajishirzi H., Farhadi A. *Editing Models with
   Task Arithmetic.* ICLR 2023. arXiv:2212.04089.
4. Zheng C., Wang Z., Ji H., Huang M., Peng N. *Model Extrapolation Expedites Alignment* (ExPO). ACL 2025.
   arXiv:2404.16792.
5. Yang E., Shen L., Guo G., Wang X., Cao X., Zhang J., Tao D. *Model Merging in LLMs, MLLMs, and Beyond: Methods,
   Theories, Applications and Opportunities.* arXiv:2408.07666, 2024. Companion list:
   https://github.com/EnnengYang/Awesome-Model-Merging-Methods-Theories-Applications
6. Jeong W., Lee W., Yoon K.-J. *Preference-Aligned LoRA Merging: Preserving Subspace Coverage and Addressing
   Directional Anisotropy.* CVPR 2026. arXiv:2603.26299.
7. Roy P. R., Roy B., Schneider K. A., Roy C. K. *A Unified Model for Cross-Domain Clone Detection via Model Merging.*
   ASE 2026. arXiv:2608.04215.
8. Chavan A., Sareen K., Tiwari R., Chandra A., Obando-Ceron J., Keutzer K., Niekum S., Dhillon I. S., Khatri D.
   *Towards Continual Learning: Mitigating Forgetting With Subspace Tuning.* arXiv:2610.22293, 2026.
9. Li Y., Yi X., Shi D., Cui Y., de Melo G., Wang L. *From Construction to Injection: Edit-Based Fingerprints for
   Large Language Models.* arXiv:2509.03122, 2025.

Learning over weights (metanetworks):

10. Zhou A., Yang K., Jiang Y., Burns K., Xu W., Sokota S., Kolter J. Z., Finn C. *Neural Functional Transformers.*
    NeurIPS 2023. arXiv:2305.13546.
11. Wolanowsky Y., Bar-Shalom G., Frasca F., Maron H. *Graph Metanetworks for Transformer Based Language Models.*
    arXiv:2610.18471, 2026.
12. Dayan A., Eitan Y., Maron H. *CrossGMN: Graph Metanetworks for Cross-Architecture Weight-Space Transformations.*
    arXiv:2610.01649, 2026.

Moving knowledge across tokenizers:

13. Minixhofer B., Vulić I., Ponti E. M. *Universal Cross-Tokenizer Distillation via Approximate Likelihood
    Matching.* NeurIPS 2025. arXiv:2503.20083.
14. Chen K., Lin C., Sun Y., Zhang W. *Dual-Vocabulary Language Model for Cross-Tokenizer Distillation.*
    arXiv:2609.33816, 2026.
15. Liu X., Yu W., Fredrikson M., Wang X., Gao J. *When the Same Coefficients Reach Different Places: Asymmetric
    Realizability in Transplanting Tokenizers across Large Language Models.* arXiv:2601.00065, 2025. First posted as
    *The Trojan in the Vocabulary: Stealthy Sabotage of LLM Composition.*

Measurement:

16. Miller E. *Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations.* arXiv:2411.00640,
    2024.

## Configuration fix

KOS-V5-Instruct also shipped with a configuration defect. Its `config.json` stores the RoPE base only inside
`rope_parameters`. vLLM 0.11 with transformers 4.57 reads the top-level `rope_theta`, finds none, and silently serves
the model at ΞΈ = 10,000. The model was trained at ΞΈ = 25,000. We confirmed this by inspecting the rotary layer vLLM
builds for each file.

V5.1 adds the top-level `"rope_theta": 25000.0` and sets `use_cache: true`. Under vLLM this restores long context,
shown in the RULER table below. Engines that read `rope_parameters` already ran KOS-V5-Instruct at 25,000, so the fix
changes nothing for them.

## Results against KOS-V5-Instruct

All comparisons below are paired runs on the same harnesses, measured on 2026-11-16. "V5" is KOS-V5-Instruct as
published, and changes are in percentage points.

Each benchmark uses its official package or the authors' released code. The only additions are the plumbing needed to
reach a locally served model. For the AfriMed-QA paper question set (n = 3,910, the paper's Table 3), one later-added
line of the released code that narrows the set to 160 questions is disabled.

### Identity (in-house probe, 15 held-out questions, greedy)

| | V5 | V5.1 |
| :--- | ---: | ---: |
| Answers as KOS (University of Kentucky and University of Louisville) | 0 / 15 | **14 / 15** |
| Claims another vendor made it (V5: almost always OpenAI) | 6 / 15 | **1 / 15** |

V5, asked "Which company developed you?", answered "I am an AI language model developed by OpenAI." The cause was the
V5 fine-tuning data. It contained the line "as an AI developed by OpenAI" about 1,800 times and never taught the model
its own identity.

### Decision battery

| Benchmark | V5 | Change | p |
| :--- | ---: | ---: | ---: |
| Med-HALT FCT (authors' code) | 8.1 | **+3.62** | 9e-106 |
| BFCL live_irrelevance (bfcl-eval, unedited) | 43.1 | **+9.62** | 2e-13 |
| Med-HALT NOTA | 60.8 | **+0.67** | 1e-5 |
| BFCL irrelevance | 64.6 | **+7.08** | 2e-5 |
| BFCL multi_turn_base | 1.0 | **+3.50** | 0.016 |
| TruthfulQA MC2 (lm-eval) | 43.4 | **+0.31** | 0.016 |
| BFCL parallel_multiple | 70.5 | **+5.00** | 0.041 |
| IFEval, four metrics (lm-eval, chat template) | 66.7 prompt-strict | βˆ’1.8 to βˆ’2.2 | β‰₯ 0.085 |
| Med-HALT FQT | 27.8 | βˆ’0.38 | 0.37 |
| AfriMed-QA MCQ, paper set (authors' code) | 32.7 | +0.46 | 0.59 |
| Other 12 BFCL categories, TruthfulQA MC1 | β€” | no significant change | β‰₯ 0.12 |

### Reasoning and knowledge

| Benchmark | V5 | V5.1 | Change | p |
| :--- | ---: | ---: | ---: | ---: |
| BBH, 27 chain-of-thought tasks, 3-shot (lm-eval) | 40.52 | 42.60 | **+2.09** | 0.0004 |
| MMLU-Pro, 14 subjects, 5-shot (lm-eval) | 29.36 | 30.23 | **+0.87** | 0.029 |
| ASDiv chain-of-thought, 8-shot (lm-eval) | 38.3 | 41.0 | **+2.78** | 3e-5 |
| HellaSwag, 10-shot | 46.7 | 46.9 | +0.21 | 0.011 |
| MMLU, 5-shot | 46.9 | 46.7 | βˆ’0.14 | 0.31 |
| GSM8K, 5-shot, strict | 52.1 | 51.6 | βˆ’0.45 | 0.62 |
| MathQA, LSAT-AR, ARC, BoolQ, PIQA, WinoGrande, OpenBookQA, COPA, CommonsenseQA, LAMBADA | β€” | β€” | no significant change | β‰₯ 0.05 |

On BBH, object counting rose +21.6, geometric shapes +18.4, multistep arithmetic +14.4 and formal fallacies +10.0.
Web of lies fell βˆ’10.0.

### Long context (RULER, 12 stock lm-eval tasks, vLLM)

| Context | V5 | V5.1 |
| :--- | ---: | ---: |
| 4K | 77.0 | **82.2** |
| 8K | 61.9 | **71.5** |
| 16K | 21.6 | **56.0** |
| 24,576 (trained context) | 11.0 | **31.1** |

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
the configuration fix. The weight program moves RULER by at most Β±2.6 points. Nothing is claimed beyond 24,576 tokens.

## Known regressions and limitations

> Disclosed deliberately.

- **Short-context RAG under noise (RGB, authors' code).**
  - **Losses:** against KOS-V5-Instruct served at ΞΈ = 10,000, noise robustness is βˆ’6.7 to βˆ’11.3 and information
    integration is βˆ’11.0.
  - **Gain:** negative rejection is +12.7.
  - **Cause:** the cost comes from the corrected RoPE base, not the weight program. The same weights at ΞΈ = 10,000
    match V5 on RGB.
  - **Rejected fix:** a frequency-hybrid RoPE configuration recovered information integration but cost Med-HALT NOTA
    βˆ’6.3, so it was rejected.
- **Three RULER tasks** lose with the configuration fix: SQuAD-QA at 4K (βˆ’6.1), multikey-3 at 8K (βˆ’15.8) and CWE at
  8K (βˆ’9.5).
- **IFEval** is 1.8 to 2.2 points lower than V5. The difference is not significant.
- **Not a knowledge model.** MMLU is 46.7. Do not use it as a medical question-answering authority.
- **MedXpertQA** is at chance level for both V5 and V5.1 (10.3 β†’ 10.9, not significant).

## Optional setting: repetition penalty 1.05

Adding `"repetition_penalty": 1.05` to `generation_config.json` also passed the gate, with 6 metrics better and none
worse. The pre-registered tie-break (count of significant gains) kept the shipped default. Against V5, this setting
gives:

- BFCL live_irrelevance +15.2 and irrelevance +9.6.
- IFEval βˆ’0.1 to +0.7.
- RGB noise robustness 4 to 6 points above V5.1's default.
- RGB negative rejection 5.7 points below V5.1's default.

Only engines that read the generation config apply it: vLLM's OpenAI server and transformers `generate`.

## Core specifications

| Attribute | Detail |
| :--- | :--- |
| **Architecture** | Decoder-only Transformer (`Qwen3ForCausalLM`), Grouped-Query Attention |
| **Parameters** | 3.715 B |
| **Hidden / Layers** | 2560 / 36 |
| **Attention** | 32 query / 8 KV heads (GQA 4:1), head_dim 128, per-head QK-RMSNorm |
| **Feed-forward** | SwiGLU, intermediate 9728 |
| **Vocabulary** | 32,000, custom medical byte-level BPE (unchanged from V5) |
| **Context length** | 24,576 trained, 32,768 configured |
| **Position encoding** | RoPE, ΞΈ = 25,000 (now in both `rope_theta` and `rope_parameters`) |
| **Embeddings** | tied |
| **Precision** | bfloat16 (7.43 GB, single shard) |

## Prompt / chat format (ChatML)

The chat template and tokenizer are byte-identical to KOS-V5-Instruct.

```
<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Write a haiku about Kentucky. Do not use any commas.<|im_end|>
<|im_start|>assistant
```

Tool / function calling uses the model's native `<tools>` … `</tools>` schema block and `<tool_call>` … `</tool_call>`
response format. Pass your function schemas via the tokenizer's `apply_chat_template(..., tools=[...])`.

## Quickstart

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Kentucky-Open-Science/KOS-V5.1-Instruct"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto")

messages = [{"role": "user", "content": "Write a haiku about Kentucky. Do not use any commas."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
```

vLLM: `vllm serve Kentucky-Open-Science/KOS-V5.1-Instruct --max-model-len 24576`. V5.1 runs at ΞΈ = 25,000 under vLLM
with no overrides.

## Related models

- [**KOS-V5-Instruct**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Instruct): the model this release reprograms.
- [**KOS-V5-Base**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Base): the from-scratch pretrained foundation
  (3.72B, 235.2B tokens).
- [**KOS-V5-Retriever**](https://huggingface.co/Kentucky-Open-Science/KOS-V5-Retriever): a LoRA retrieval adapter
  built for KOS-V5-Instruct. It has not been tested with V5.1.

## Intended use & limitations

Research use only. English only. Not for clinical, commercial, legal, or production-grade use. Outputs may be
factually wrong or fabricated. This model must not be used to make or inform medical decisions.

## Naming

The program is **KOS** (KOS-V1..V6). V5.1 is a weight-space-programmed release within the KOS-V5 generation.