Vertex-0.6-100M-8192-Instruct-v2

Vertex 0.6 100M — 8192-ctx Instruct v2

Instruction-tuned chat model of the Vertex 0.6 family by Vertex Research, and the most capable chat model in the family so far. Built on a quality-annealed 8192-context base, then SFT'd on ~201K conversations weighted toward multi-turn dialogue so it holds a conversation across several turns instead of unravelling after the first message.

Model details

Usage

from transformers import AutoTokenizer, AutoModelForCausalLM

repo = "VertexResearch/Vertex-0.6-100M-8192-Instruct-v2"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)

msgs = [{"role": "user", "content": "Who are you?"}]
text = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False)
ids = tok(text, return_tensors="pt", add_special_tokens=False).input_ids
out = model.generate(ids, max_new_tokens=120, eos_token_id=[2, 6])
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

For tool calling, put function definitions in the system prompt; the model emits <tool_call>{"name": ..., "arguments": ...}</tool_call> and consumes results in <tool_response> blocks.

Training

Base: an 8192-context Vertex 0.6 100M base given a 1B-token quality anneal (FineWeb-Edu + synthetic elementary word-problem math). SFT with TRL on ~201K conversations: smol-smoltalk (~130K, concise multi-turn), UltraChat 200k (~40K, longer multi-turn), everyday-conversations (basic-chat grounding, 2×), function/tool-calling data, QA/tutoring, and self-identity. 2 epochs, lr 3e-4 cosine, bf16, max length 2048. Final eval loss 1.404.

Limitations

These models are not the most coherent yet and need more tuning: expect rambling, repetition, and inconsistent answers, especially over longer generations.

97M parameters: multi-turn chat is much improved and tool-call syntax works, but factual accuracy is low, reasoning is shallow, and it makes arithmetic errors. Not for production use. Knowledge cutoff ~April 2024.

Tool calling works mechanically (correct <tool_call> format, stops cleanly) with a single available function, but multi-step tool use is unreliable: with more than one function available it can pick the wrong tool, and it can hallucinate details when summarizing a tool's response rather than reporting it accurately. Don't trust it in an unsupervised agent loop.