Boris-2-Think: Difference between revisions
Appearance
New article with Infobox model (data from the Hugging Face model card) |
Release date in Pacific time (2026-11-11 10:44 PM) |
||
| Line 45: | Line 45: | ||
| formats = safetensors (transformers, trust_remote_code) | | formats = safetensors (transformers, trust_remote_code) | ||
| website = [https://oc.kunix.org oc.kunix.org] | | website = [https://oc.kunix.org oc.kunix.org] | ||
| released = 2026-11- | | released = 2026-11-11 | ||
}} | }} | ||
'''Boris-2-Think''' is [[Boris-2]] fine-tuned to write a short reasoning trace inside <code><think></code> tags before a <code>Final answer:</code> line. It was trained on [[OpenThink]], OpenCerebral's set of 43k short reasoning traces written for models this size. | '''Boris-2-Think''' is [[Boris-2]] fine-tuned to write a short reasoning trace inside <code><think></code> tags before a <code>Final answer:</code> line. It was trained on [[OpenThink]], OpenCerebral's set of 43k short reasoning traces written for models this size. | ||
Latest revision as of 00:50, 16 November 2026
Boris-2-Think
Overview
Developer
Family
Boris-2
Type
Reasoning (question answering)
Fine-tuned from
License
Apache 2.0
Status
Released
Architecture
Architecture
Decoder-only transformer: 8-head MHA, SwiGLU 1664, RoPE, Canon layers, 15M n-gram table
Parameters
143.7M
Layers
31 (dmodel 512)
Context length
2,048 tokens
Tokenizer
Own byte-level BPE, 32,768; ChatML-style turns as plain text
Training
Tokens
200.04B (via Boris-2)
Tokens / param
1,392 Chinchilla-optimal ≈ 20
Data
Fine-tune: OpenThink, hybrid recipe (42,301 rows, 2 epochs)
Hardware
1 × NVIDIA V100 32GB
Training time
About 5 minutes (fine-tune, 404 steps)
Benchmarks
Benchmark
Score
Boris-2
HellaSwag
50.13
50.27
ARC-Easy
63.93
59.76
ARC-Challenge
34.81
32.00
PIQA
72.31
72.85
WinoGrande
53.67
52.57
OpenBookQA
34.40
36.60
SciQ
87.10
85.10
MMLU (continuation)
29.30
28.96
CommonsenseQA
20.15
20.88
ArithMark-3
49.20
45.10
Mean 9 tasks
49.5
48.8
Zero-shot, lm-evaluation-harness 0.4.12, acc_norm (WinoGrande: acc). Nine-task mean; ArithMark-3 is left out · Source: benchmarks_comparison.csv
Links
Weights
Formats
safetensors (transformers, trust_remote_code)
Website
- Released11 days ago
Boris-2-Think is Boris-2 fine-tuned to write a short reasoning trace inside <think> tags before a Final answer: line. It was trained on OpenThink, OpenCerebral's set of 43k short reasoning traces written for models this size.
External links
Image by Jakub Steiner / Wikimedia Commons / Licensed under CC BY-SA 2.0
This page or section is a stub. You can help the wiki by accurately contributing to it.