---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
language: en
base_model: allenai/Olmo-3-1025-7B
tags:
- continued-pretraining
- merged
- bfloat16
---

# OLMo-3-7B-CPT-BF16

Standalone **BF16 (16-bit)** continued-pretraining model. This revision contains
the complete model from **epoch 2, step 750**, in exactly **eight
safetensors weight shards**. Load directly with Transformers.

## Epochs

`main` contains epoch 2, the best training-time validation checkpoint. Each
completed epoch is independently loadable using its `epoch-N` revision. Both
training runs stopped after four epochs; there is no epoch-five export.

| Revision | Training step | Training-time validation loss |
|---|---:|---:|
| [epoch-1](https://huggingface.co/CompassioninMachineLearning/OLMo-3-7B-CPT-BF16/tree/epoch-1) | 375 | 1.3267499 |
| [epoch-2](https://huggingface.co/CompassioninMachineLearning/OLMo-3-7B-CPT-BF16/tree/epoch-2) | 750 | 1.3001196 |
| [epoch-3](https://huggingface.co/CompassioninMachineLearning/OLMo-3-7B-CPT-BF16/tree/epoch-3) | 1125 | 1.3246491 |
| [epoch-4](https://huggingface.co/CompassioninMachineLearning/OLMo-3-7B-CPT-BF16/tree/epoch-4) | 1500 | 1.3742925 |

These losses were measured during adapter training, not by a new evaluation of
the BF16 exports. Evaluate the four revisions separately to measure behavior
changes across epochs.

## Load a specific epoch

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "CompassioninMachineLearning/OLMo-3-7B-CPT-BF16"
revision = "epoch-2"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
model = AutoModelForCausalLM.from_pretrained(
    model_id, revision=revision, dtype=torch.bfloat16, device_map="auto"
)
```

## Merge provenance and validation

- Original base: `allenai/Olmo-3-1025-7B` at `a81bae42db3975be1671e27b9c9a56da1a9f980f`.
- Source adapter: [checkpoint 750](https://huggingface.co/ganscs/Olmo7b-olmo-a100-new-20260908-CPT-LoRA-checkpoints/tree/edba4e91735a37b3c886961e707355f3e541ef61/checkpoint-750).
- Adapter repository revision: `edba4e91735a37b3c886961e707355f3e541ef61`.
- Adapter SHA-256: `c1486c306d6a1a9b5a77d6abd8151f0a4f0f4a4051397181db754dda4700f04d`.
- Merged 224 rsLoRA layers with PEFT's safe merge.
- The separately trained `embed_tokens` and `lm_head` matrices are included.

Training used a 4-bit base; these exports merge the trained adapter into the
pinned original BF16 base. The merge was performed using PEFT's safe merge,
not Unsloth's export helper. Every tensor was checked for finite BF16 values and
the complete architecture's names and shapes. Every epoch passed a standalone
load, finite-logit forward pass and short greedy generation. No separate adapter
is required, and optimizer/trainer checkpoint files are not included.

`merge_manifest.json` records source and output checksums. `validation.json`
records the inference smoke test. `training_manifest.json` preserves the training
settings. Shard count controls packaging; BF16 specifies the 16-bit precision.