wgaca/Spomin-Qwen2.5-3B-Summary-v2-GGUF

Spomin Summary Worker v2 — Q8_0 GGUF

Built with Qwen. Modified weights: this is a Spomin fine-tune of Qwen2.5-3B-Instruct, merged from the selected epoch-2 rank-32 adapter and converted to Q8_0 GGUF. It is not the unmodified Qwen model and is not a general-purpose chat assistant.

This worker summarizes one immutable conversation/source chunk for the Spomin router. It returns a bounded JSON summary. The router owns chunk selection, validation, exact source preservation, summary installation, and surgical KV edits; the worker does not perform those operations or decide what the router should retain.

Download and identity

hf download wgaca/Spomin-Qwen2.5-3B-Summary-v2-GGUF Spomin-Qwen2.5-3B-Summary-v2-Q8_0.gguf --local-dir .
Get-FileHash .\Spomin-Qwen2.5-3B-Summary-v2-Q8_0.gguf -Algorithm SHA256

The included artifact manifest identifies this release. This repository contains the deployed Q8_0 worker, not the larger BF16 export, adapters, training data, private transcripts, or local evaluation logs.

Serve it for Spomin

With a compatible llama-server build:

llama-server -m .\Spomin-Qwen2.5-3B-Summary-v2-Q8_0.gguf --alias qwen2.5-3b-spomin-summary-v2 --host 127.0.0.1 --port 8088 --ctx-size 16384 --parallel 1 -ngl 999

Adjust GPU placement for your machine. Do not replace or interrupt the primary model runtime to start the worker. In Spomin, configure this matched pair:

{
  "worker_backend_url": "http://127.0.0.1:8088/v1",
  "worker_model_id": "qwen2.5-3b-spomin-summary-v2",
  "worker_contract_version": "spomin.summary-only-contract.v2",
  "worker_backend_max_completion_tokens": 512
}

The worker model service is separate from the primary runtime. Managed Spomin requires the KV-surgical llama.cpp fork for its primary slot. The harness connects to the router and must disable its own automatic and manual compaction when using server-managed context.

Worker contract

Use the exact system prompt and input schema. The user payload has task SUMMARIZE_SOURCE and one source containing its identity, observed role, token count, summary limits, and content. No neighboring conversation, FACTS, GIST, retention decisions, or rolling state is supplied to this worker.

The output schema is:

{"schema_version":"spomin.summary-output.v2","summary":"..."}

Respect the requested per-source token and word limits and require compression relative to the original source. Validate output rather than trusting syntactically plausible JSON. The v2 input schema permits ceilings up to 384 tokens / 240 words; a request can impose smaller limits. Inputs below 300 source tokens are passed through by the router; sources above 14,000 tokens are split at tokenizer boundaries before independent worker calls.

Validation and limitations

This is the artifact selected for Spomin's v2 worker deployment. Its selection used validation loss and a bounded ten-example generation comparison. Local checks covered schema-valid output and a managed summary installation with exact source preservation. Those are limited mechanical checks, not a broad guarantee of semantic fidelity. The exhaustive semantic, retrieval, regression, and downstream-primary evaluation gates were not completed under that selection's evaluation budget.

Summaries can omit details or misrepresent source material. Preserve the original source, validate outputs, and evaluate the model on your workload. This model card does not claim new long-workload benchmark results or unlimited native context continuity.

License and attribution

This derivative is distributed under the included Qwen Research License, not Apache-2.0 or MIT. It permits non-commercial research/evaluation use subject to its terms; commercial use requires a separate license from Alibaba Cloud. See the full agreement and Notice. No additional commercial-use permission is granted here.