bigsk1/ jarvis-helper:minicpm5-1b-q4_k_m-v3

57 2 months ago

Jarvis-specific MiniCPM5-1B Q4_K_M fine-tune for fast local status phrases and factual summaries; 98.1% validation and 96.4% protected-test accuracy. 

ollama run bigsk1/jarvis-helper:minicpm5-1b-q4_k_m-v3

Details

2 months ago

d9b7edc7004b · 688MB

model
llama
·
1.08B
·
Q4_K_M
params
{ "num_ctx": 8192, "stop": [ "<|im_end|>", "</s>" ], "temperature":
template
{{- if .Messages -}} {{- range .Messages -}} <|im_start|>{{ .Role }} {{ .Content }}<|im_end|> {{ end

Readme

Jarvis Helper V3

Jarvis Helper V3 is a Jarvis-specific LoRA fine-tune of OpenBMB MiniCPM5-1B, exported as Q4_K_M GGUF for fast local helper tasks through Ollama.

This is no longer the unchanged upstream GGUF packaged by the original V1 release. V3 has adapted weights trained for the bounded prompt contracts used by Jarvis Voice. It remains under the bigsk1/jarvis-helper namespace so it cannot be mistaken for an official OpenBMB release. OpenBMB is the upstream model developer and is not responsible for this fine-tune.

Quick start

ollama pull bigsk1/jarvis-helper:minicpm5-1b-q4_k_m-v3
ollama run bigsk1/jarvis-helper:minicpm5-1b-q4_k_m-v3

Intended jobs

V3 was trained for five small, bounded transformations:

  • 3-8 word progress/status phrases without false completion claims;
  • dense factual Stash summaries;
  • longer text summaries that preserve numbers, names, caveats, and uncertainty;
  • four-label content classification with strict JSON;
  • deterministic entity/value extraction with strict JSON.

Jarvis currently promotes V3 for status phrases, Stash summaries, and its separately opt-in text summarizer. Classification and extraction remain non-production evaluation roles. The Helper stays independent of Jarvis’s primary chat, vision, and embedding models.

Fine-tuning details

  • Base checkpoint: final post-trained openbmb/MiniCPM5-1B BF16 release
  • Method: LoRA with assistant-completion-only loss
  • Dataset: 1,300 training, 120 validation, and 120 protected-test records
  • Tasks: status, Stash summary, text summary, classification, extraction
  • Epochs: 2
  • Learning rate: 1e-4
  • LoRA rank / alpha / dropout: 16 / 32 / 0.05
  • Training context: 2,048 tokens
  • Export: merged full model, Q4_K_M GGUF
  • Runtime context configured by Jarvis: 8,192 tokens
  • GGUF SHA-256: 77a530bec3ef638101143d9420dea35e44438b18b999d61abf107d952ee3e672

The production-shaped dataset uses stable TASK=... identifiers shared with Jarvis’s runtime prompts. Validation and test inputs were isolated from training. The protected V3 test was consumed only after the candidate artifact was frozen.

Measured improvement

On the same 120 unseen V3 validation cases, using one round for a direct comparison:

Model Overall Status Stash Text Classification Extraction
Original MiniCPM5 Q4_K_M package 20.00% 0.00% 20.00% 16.67% 65.00% 0.00%
Jarvis Helper V2-final 55.83% 100.00% 30.00% 46.67% 95.00% 0.00%
Jarvis Helper V3 98.33% 100.00% 93.33% 100.00% 100.00% 100.00%

Across three validation rounds / 360 requests, V3 scored 98.06% overall with zero request errors and zero forbidden summary contradictions. Its frozen three-round protected test scored 96.39% overall:

  • status: 96.00%;
  • Stash summary: 100.00%;
  • text summary: 95.56%;
  • classification: 100.00%;
  • extraction: 86.67% exact;
  • median / p95 wall latency: 861 ms / 1,163 ms.

The standard Jarvis Helper benchmark measured 133 ms warm median latency, 94.1% summary fact retention, 100% status-format compliance, and 100% exact classification.

Size and runtime memory

The downloadable GGUF is approximately 688 MB. ollama ps may show roughly 849 MB while loaded because resident memory also includes runtime bookkeeping, compute buffers, and the KV cache. That is expected and is not additional model weight growth from fine-tuning.

Jarvis configuration

JARVIS_HELPER_LLM_BASE_URL="http://127.0.0.1:11434"
JARVIS_HELPER_LLM_MODEL="bigsk1/jarvis-helper:minicpm5-1b-q4_k_m-v3"
JARVIS_HELPER_LLM_DEVICE="auto"

STATUS_LLM_PROVIDER=helper
STASH_SUMMARIZE_LLM_PROVIDER=helper
TEXT_SUMMARIZER_LLM_PROVIDER=helper

Native Jarvis installations can pull the selected mode’s configured model with:

./bin/setup-helper-llm --mode cloud

Replace cloud with local for local mode.

Limitations

This compact text-only 1B model is not a replacement for a primary chat model. Do not use it for embeddings, vision, complex reasoning, high-stakes decisions, or autonomous tool selection.

The protected test found a repeatable extraction boundary error: V3 sometimes combines an organization and adjacent product into one entity array item. Extraction is therefore not enabled as a production Jarvis role. Applications should continue to validate structured output and retain deterministic or primary-model fallbacks.

Versioning and attribution

  • minicpm5-1b-q4_k_m-v1: original upstream Q4_K_M packaging release
  • minicpm5-1b-q4_k_m-v3: Jarvis-specific fine-tuned release

MiniCPM5 was developed by OpenBMB and released under Apache 2.0. Jarvis Helper V3 is a derivative fine-tune packaged for local Jarvis integration.

Project: github.com/bigsk1/jarvis-voice