guoxuter/ov_intent_analysis_sft

OpenViking Intent Analysis SFT v7

ov_intent_analysis_sft is a Qwen3.5-0.8B model fine-tuned for OpenViking retrieval intent analysis and query planning. Given recent conversation context and the current user message, it decides whether retrieval is needed and emits structured queries targeting OpenViking skill, resource, and memory scopes.

This repository contains the original Transformers checkpoint in Safetensors format. The corresponding quantized Ollama release is `guoxuter/ov_intent_analysis_sft:v7_q8`.

Loading

Use a recent Transformers version with Qwen3.5 support:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "guoxuter/ov_intent_analysis_sft"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

The model was trained for the OpenViking v7 retrieval prompt and structured output contract. For end-to-end use, prefer the prompt bundled with OpenViking rather than a generic chat prompt.

Artifact provenance

• —Base model: `Qwen/Qwen3.5-0.8B`
• —Training checkpoint: step 600 of the v7 SFT run
• —Safetensors SHA-256: c92c878a96f34d2f0c87d2308099de7dc1401aae58aba0310aca550e9024b33b
• —Corresponding Ollama Q8 GGUF layer SHA-256: aa98adccdec6a3be82d462563586abe1db520f93281ecc3f9bc3ff978b12d795

The Safetensors checkpoint is the source artifact. The Ollama model is a Q8 GGUF derivative and should not be used to reconstruct full-precision weights.

Intended use

This model is intended as a compact retrieval planner for OpenViking-compatible systems. It is not a general-purpose assistant. Outputs should be validated against the expected structured schema before they are executed or used for retrieval.

License

The base Qwen3.5-0.8B model is released under the Apache License 2.0. This fine-tuned checkpoint is published under the same license.