Boris-2-Preview-2

The second preview of Boris-2, a 144M-parameter base language model from OpenCerebral, trained from scratch on a single NVIDIA V100. This preview is the main run's 100B-token checkpoint with a short learning-rate decay (2.5B tokens) applied, the same decay length as Boris-2-Preview-1. The full run continues to 200B tokens.

This is a base model. It continues text; it does not follow instructions or chat.

Results

Zero-shot, on the float32 weights in this repository: lm-evaluation-harness 0.4.12 for the first four, the leaderboard's own bencharithmark-3.py (unmodified, default settings) for ArithMark-3.

The Index is the Open SLM Leaderboard formula: (N(HellaSwag,25) + N(mean(ARC-E, ARC-C),25) + N(PIQA,50) + 0.65 * N(ArithMark-3,25)) / 3.65, with N(v,c) = 100(v-c)/(100-c).

Single-checkpoint scores on these benchmarks move by roughly plus or minus 0.8 Index between neighbouring checkpoints. No checkpoint or seed was selected on these scores: the decay start (100B) and length (2.5B tokens) were fixed in advance, and this is the one checkpoint at its end. The decay data mix was chosen from 200M-token pilot decays on an earlier checkpoint, scored on these benchmarks plus SVAMP and ASDiv as held-out word-problem checks.

The decay moved the scores from HellaSwag 40.75, ARC-Easy 56.19, ARC-Challenge 29.18 and PIQA 69.37 at the 100B checkpoint to the figures above. ArithMark-3 rose 4.6 points over Preview-1, from the grade-school word-problem data added to the decay. That gain is specific to ArithMark's skills so far: on SVAMP (35.3) and ASDiv (39.3), scored as four-way multiple choice, Preview-2 is level with the pilot baseline (36.0 and 38.5).

Model

Two components differ from a standard decoder:

• Canon layers. Each block applies a causal 3-tap depthwise convolution to the fused query/key/value projection, so every position mixes in the two previous positions before attention.
• N-gram table. An exact lookup on the last two and three token ids (210,938 bigrams and 23,437 trigrams, no hashing) retrieves a learned rank-64 vector that is added to the residual stream before block 2. It is a gather, not a matrix multiply.

Training packed several documents per sequence with attention, Canon and the n-gram lookup all masked at document boundaries. The modeling code reproduces this: a new document starts after every end-of-text token (eos_resets_context in the config).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "opencerebral/Boris-2-Preview-2"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)

x = tok("The water cycle begins when", return_tensors="pt")
y = model.generate(**x, max_new_tokens=40, do_sample=False)
print(tok.decode(y[0], skip_special_tokens=True))

The custom code is plain PyTorch with no extra dependencies. It has no KV cache, so generation recomputes the prefix every step; at this size that is fast enough for sampling and evaluation. Run in float32 or bfloat16. The residual stream is kept in float32 internally because one channel reaches values near the float16 limit.

The converted model was checked against the training code on the same weights: logits match to within 4e-5 in float32, and lm-evaluation-harness reproduces the training evaluator's benchmark scores to within one item.

Training

Main run data, by tokens: up to 55B tokens, about 89% ClimbMix and 11% FineMath; from 55B tokens, about 90% ClimbMix, 10% UltraData-Math (L3 synthetic QA) and 0.15% Orca-Math.

Decay data, by tokens:

Contamination

No benchmark training, validation or test split was trained on. MetaMathQA's rows are rewrites of the GSM8K training set; GSM8K is not one of the evaluated benchmarks. ASDiv and SVAMP, used only as held-out checks during development, were not trained on.

• Every source except ClimbMix was filtered against every evaluation item (HellaSwag, ARC, PIQA, ArithMark-3) with a 13-gram overlap check. ClimbMix, the bulk of the data, has not been filtered.
• About two thirds of HellaSwag is built from WikiHow articles, and the Cosmopedia tutorials were generated from WikiHow titles. As for Preview-1, the validation items were split by whether their WikiHow title is also one of the 177,353 tutorial titles, and the 100B checkpoint was compared with this one:

The two WikiHow groups gained the same amount (0.2 points apart, well within noise). The gain is a model getting better at how-to text in general, not at the specific topics in the tutorial data.

Limitations

A 144M-parameter base model trained on 102.5B tokens. It makes factual and arithmetic mistakes, loses track of longer arguments, and repeats itself under greedy decoding. It has no safety tuning. It is a research preview, not a product.

Sample outputs

Greedy decoding, float32, unedited, cut for length. The same twelve prompts as Preview-1; all twelve are in samples.json.

Compared with Preview-1, the train question and France's capital now come out right, and the last prompt no longer loops. 47 + 68 is still wrong (105), the egg is cracked before boiling, and the shoelace prompt drifts into sewing. The word-problem data shows in the answer format (#### 15, The answer is:).