Update family README (checksums table)
Browse files
README.md
CHANGED
|
@@ -23,6 +23,41 @@ project — open source, open data, open weights.
|
|
| 23 |
| [ANDREA-12M](#andrea-12m) | 12.8M | 384d / 12h / 6L / 1024ctx | Shipped 2026-03-21 |
|
| 24 |
| [ANDREA-120M](#andrea-120m) | 98.7M | 768d / 12h / 12L / 1024ctx | Shipped 2026-05-27 |
|
| 25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
---
|
| 27 |
|
| 28 |
# ANDREA-12M
|
|
|
|
| 23 |
| [ANDREA-12M](#andrea-12m) | 12.8M | 384d / 12h / 6L / 1024ctx | Shipped 2026-03-21 |
|
| 24 |
| [ANDREA-120M](#andrea-120m) | 98.7M | 768d / 12h / 12L / 1024ctx | Shipped 2026-05-27 |
|
| 25 |
|
| 26 |
+
## Training artifacts here. Deployment weights next door.
|
| 27 |
+
|
| 28 |
+
**This repo holds our training side.** Raw CUDA checkpoints (`.bin`), which
|
| 29 |
+
carry Adam's first and second moments alongside the weights so a run can be
|
| 30 |
+
resumed mid-flight, plus portable JSON for our own pure-Python engine, bandit
|
| 31 |
+
state, and loss history. This is what you want to reproduce or continue
|
| 32 |
+
training. It is not what you want to run — `transformers` cannot load any of it.
|
| 33 |
+
|
| 34 |
+
**Deployment lives in the per-model repos**: safetensors, config, tokenizer,
|
| 35 |
+
and a reference implementation, ready for `from_pretrained`.
|
| 36 |
+
|
| 37 |
+
| Model | Deployment repo | safetensors |
|
| 38 |
+
|-------|-----------------|-------------|
|
| 39 |
+
| ANDREA-120M | [russellbal/ANDREA-120M](https://huggingface.co/russellbal/ANDREA-120M) | 376.5 MB |
|
| 40 |
+
| ANDREA-12M | [russellbal/ANDREA-12M](https://huggingface.co/russellbal/ANDREA-12M) | 48.8 MB |
|
| 41 |
+
|
| 42 |
+
```python
|
| 43 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 44 |
+
|
| 45 |
+
tok = AutoTokenizer.from_pretrained("russellbal/ANDREA-120M", trust_remote_code=True)
|
| 46 |
+
model = AutoModelForCausalLM.from_pretrained("russellbal/ANDREA-120M", trust_remote_code=True)
|
| 47 |
+
|
| 48 |
+
ids = tok.apply_chat_template(
|
| 49 |
+
[{"role": "user", "content": "What is a permacomputer?"}],
|
| 50 |
+
add_generation_prompt=True, tokenize=False)
|
| 51 |
+
out = model.generate(tok(ids, return_tensors="pt").input_ids,
|
| 52 |
+
max_new_tokens=64, do_sample=True, temperature=0.7)
|
| 53 |
+
print(tok.decode(out[0]))
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
The deployment weights are the same numbers as the training checkpoints here:
|
| 57 |
+
the export drops the optimizer state and is verified tensor by tensor against
|
| 58 |
+
its source `.bin`, and the published files are sha256-matched against that
|
| 59 |
+
verified build.
|
| 60 |
+
|
| 61 |
---
|
| 62 |
|
| 63 |
# ANDREA-12M
|