File size: 1,044 Bytes
c1c30aa
0599433
 
c1c30aa
ba5d0c1
c1c30aa
 
ba5d0c1
c1c30aa
ba5d0c1
 
 
7ed6d30
c1c30aa
 
7a9c479
c1c30aa
ba5d0c1
c1c30aa
ba5d0c1
 
 
c1c30aa
ba5d0c1
c1c30aa
ba5d0c1
 
 
c1c30aa
ba5d0c1
 
 
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
---
base_model:
- allenai/Olmo-3-7B-Instruct
library_name: transformers
model_name: Role-mo-V2-32B
tags:
- dpo
- trl
licence: license
datasets:
- nvidia/HelpSteer3
- ConicCat/Lamp-P-Prompted
- ConicCat/C2-Nemo-Delta-RLCD-Preference
---

# Model Card for ConicCat/Role-mo-V2-7B

This model is a fine-tuned version of [allenai/Olmo-3.1-32B-Instruct](https://huggingface.co/allenai/Olmo-3.1-32B-Instruct) using humanline dpo to improve writing, roleplay, and chat capabilities.

## Sampler Settings
* Chatml template
* I recommend .7 temp and (optionally) 1.05 rep pen.

## Datasets

* nvidia/HelpSteer3 to maintain general capabilites and improve chat performance.
* ConicCat/Lamp-P-Prompted for improved prose and slop reduction.
* A C2 based human prompt / synthetic response roleplay preference dataset.

## Changelog
* V0 > V1: Changed preference data construction to improve margins between the chosen and rejected responses.
* V1 > V2: More optimal hyperparameters; lowering lr and increasing bsz allows for stable training with lower beta.