---
license: apache-2.0
pipeline_tag: text-generation
library_name: transformers
language:
- en
tags:
- from-scratch
- experimental
- causal-lm
- small-language-model
- instruct-pretrained
datasets:
- HuggingFaceH4/ultrachat_200k
- allenai/soda
- openbmb/UltraInteract_sft
- microsoft/orca-math-word-problems-200k
- databricks/databricks-dolly-15k
- b-mc2/sql-create-context
---

# Ivme-Conversate-S-v2-Instruct

9,021,600 parameters. Standard decoder-only Transformer (tied
embeddings, multi-head attention, RoPE, SwiGLU, RMSNorm) -- matching
Ivme-Conversate-v2-Base's proven recipe exactly, deliberately with zero
architectural novelty.

Trained single-epoch on ~900M tokens, instruct-heavy from the start rather
than base-pretrain-then-finetune: UltraChat-200k (real multi-turn dialogue)
as the dominant 45% share, plus SODA, UltraInteract reasoning traces,
orca-math, dolly-15k instructions, and sql-create-context. All sources
permissively licensed (MIT/CC-BY/CC-BY-SA).

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "ivmelabs/Ivme-Conversate-S-v2-Instruct", trust_remote_code=True
)
tok = AutoTokenizer.from_pretrained("ivmelabs/Ivme-Conversate-S-v2-Instruct")

ids = tok("Hello!", return_tensors="pt").input_ids
out = model.generate(ids, max_new_tokens=80, do_sample=True, temperature=0.8, top_k=40)
print(tok.decode(out[0]))
```

Note: no KV-cache in this architecture -- `.generate()` works but is O(n^2)
rather than O(n), fine for short samples, not tuned for long-form serving.