File size: 4,339 Bytes
1c016ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9b4717b
1c016ea
 
9b4717b
 
a10c7e3
1c016ea
 
 
 
 
 
 
 
9b4717b
 
 
 
 
1c016ea
9b4717b
 
 
 
 
1c016ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1ac955a
1c016ea
 
 
 
 
 
 
 
1ac955a
1c016ea
1ac955a
1c016ea
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9b4717b
1c016ea
 
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
---
library_name: transformers
license: apache-2.0
tags:
- oleander
- gdn2
- gated-deltanet
- gqa
- muon
- marimo
- molab
---

# Oleander-150M

**Training checkpoint** (`round 14`) β€” step **7,000 / 28,500** Β· **24.56%** of the planned run.

```text
steps    β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘  24.56%   7,000 / 28,500
tokens   β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘β–‘  19.30%   2.88B / 14.94B compute
phase    tst   (TST bag training ends at step 9,000)
```

Trained with the Oleander marimo trainer (molab): hybrid **GDN2** (gated delta-rule, gated-delta-net-2) and **gated GQA** layers with **block attention residuals**, MuonEq-R + AdamW, WSD schedule.

## Training progress

| | |
|---|---|
| Step | **7,000** / 28,500 |
| % complete (steps) | **24.56%** |
| Compute tokens seen | **2.88B** / 14.94B (19.30%) |
| Raw tokens seen (TST bags unpacked) | 17.30B / 38.54B |
| Cluster tokens (all 1 worker(s)) | 2.88B / 14.94B compute |
| Phase | `tst` |
| Val loss / ppl | **6.3489** / **571.85** |
| Train loss (latest) | 5.2985 |
| Throughput | 76,860 compute tok/s Β· 0.139 steps/s |
| Session elapsed | 0d 03h 03m Β· ETA 1d 18h 57m |
| Checkpoint written | 2026-11-15T09:39:11Z |

## How to load

The repo is self-contained: `config.json` carries an `auto_map` that points at `configuration_oleander.py` and `modeling_oleander.py` (shipped here), so remote code is all you need.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "CyanMonkey/Test-Subject-3"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)

prompt = "Once upon a time"
out = model.generate(**tok(prompt, return_tensors="pt"), max_new_tokens=64,
                     do_sample=True, temperature=0.8, top_k=50)
print(tok.decode(out[0], skip_special_tokens=True))
```

The GDN2 mixers use `flash-linear-attention` kernels when they are installed and you run in bf16 on a GPU (`pip install --no-deps fla-core flash-linear-attention`); otherwise a pure-PyTorch recurrent reference is used β€” the same math, but slower for long prompts. There is no KV cache (`use_cache=False`).

## Architecture

```text
Oleander-150M (28L = 7 x [GDN2 Β· GDN2 Β· GDN2 Β· GQA] Β· d=576 Β· SwiGLU(1600) Β· tied 32,770 vocab)
```

| | |
|---|---|
| Parameters | 149,976,139 (149.98 M) |
| Layers | 28 = 7 x [gdn2, gdn2, gdn2, gqa] |
| Hidden size / FFN | 576 / 1600 (SwiGLU) |
| Vocab / context | 32,770 (tied embeddings) / 1,024 |
| Features | GDN2 (gated delta net 2) x21: 9h x 64, conv4, gated GQA x7: 9 Q / 3 KV heads, head_dim 64, QK-Norm (offset form), partial RoPE (rotary_dim 32, theta 100000), value residual (v_mix), SwiGLU, RMSNorm, AttnRes (block attention residuals), tied embeddings, TST (token superposition training, bag 6) |

## Training setup

| | |
|---|---|
| Optimizer | muoneq_r (Muon) + AdamW β€” `muon_adamw` |
| LR schedule | WSD Β· warmup 100 Β· stable until 82.5% Β· 1-sqrt cooldown |
| Batch | 524,288 compute tokens / step (32 x 16 x 1024) |
| TST | True Β· bag 6 Β· first 30% of steps |
| Data | `CyanMonkey/NextGen-1-32-1024` (train), 14,753,933 rows x 1024 tokens |
| Precision | bf16 autocast Β· weights exported as float32 |

## Files

| file | what |
|---|---|
| `model.safetensors` | weights, trainer parameter names |
| `config.json` | `auto_map` -> `AutoConfig` / `AutoModelForCausalLM` remote-code entry points |
| `configuration_oleander.py` | `OleanderConfig` |
| `modeling_oleander.py` | `OleanderForCausalLM` (same math as the trainer) |
| `tokenizer.json`, `tokenizer_config.json`, `special_tokens_map.json` | tokenizer |
| `training_meta.json` | detailed run metadata (this checkpoint) |
| `progress.json` | flat step / tokens / % summary |
| `generation_config.json`, `train_config.json` | sampling defaults, full trainer config |

## Notes

- This is a **work-in-progress** checkpoint at 24.56% of the planned run, not a finished model: expect repetition and incoherence, especially with few-shot prompts.
- The tokenizer comes from the base model the run started from; the vocabulary is a plain BPE model and is not extended during training.
- `training_meta.json` is regenerated on every push and is the authoritative record of how far training got.