AuraGo-Spark

Q4_K_M GGUF of Spark-X2.5-4B, fine-tuned for AuraGo tool usage and framework knowledge, including the first targeted correction round.

This is an experimental model with known language and instruction-following limitations. It is a language model used by AuraGo, not the complete AuraGo application.

Download

File	Quantization	Size
AuraGo-Spark-X2.5-4B-Q4_K_M.gguf	Q4_K_M	2,600,224,352 bytes

SHA256:

3683da863f81a0f2f6c752fe47447845bf83f09fa7548d6f62407846f163039b

The file is a direct BF16-to-Q4_K_M quantization of the merged fine-tune, without requantization, pruning or an importance matrix. All 290 tensors are retained: 180 Q4_K, 37 Q6_K and 73 F32.

Compatible engine

Use a Spark-capable version of llama-wackMall-hybrid. Conversion and local verification used commit dacde9fd2414ebc0533d53353498bdd5409a670f. Older Ling-only or generic llama.cpp builds may not support spark2_5.

The original Spark tokenizer and chat template are embedded. Use thinking off, one slot, speculative decoding off and KVFlash off. Do not substitute a Qwen or Ling template.

Example with an existing compatible server and a separately created API-key file:

hf download antibyte/AuraGo-Spark AuraGo-Spark-X2.5-4B-Q4_K_M.gguf --local-dir ./AuraGo-Spark

llama-server \
  -m ./AuraGo-Spark/AuraGo-Spark-X2.5-4B-Q4_K_M.gguf \
  --host 127.0.0.1 --port 8080 --api-key-file api-key.txt \
  -c 4096 -np 1 --reasoning off --spec-type none

Use the fork's hardware-specific launch settings for GPU execution. The local export smoke test used a 4096-token context on a Windows CPU build. Larger contexts and GPU performance were not validated for this fine-tuned GGUF. Reported speeds for the original Spark model are not measurements of this export.

Training provenance

• Base: XHToken/Spark-X2.5-4B, pinned revision 5e10fcc0286756aebf7c41dc52c1e42d95c70281.
• Architecture: 4.112B parameters, 36 layers, hybrid sliding/full attention, sliding window 512, head-wise attention gates.
• Initial supervised fine-tune: 922 reviewed synthetic AuraGo dialogs, three epochs, LoRA rank 16, alpha 32, dropout 0.05, learning rate 5e-5, effective batch 8. Targets: q_k_v_proj, g_proj, out_proj, gate_proj, up_proj, down_proj.
• Correction 1: continued the initial adapter with a fresh optimizer, 216 dialogs, two short epochs, learning rate 1e-5, effective batch 8; selected checkpoint 54. The mix includes handwritten language examples, deterministic arithmetic, tool-operation contrasts and 96 original train-only replay dialogs.
• The original Spark template was used with thinking disabled. Training supervised assistant bodies and end-of-sequence tokens; other roles and headers were masked.
• Training ran on one NVIDIA A100 SXM 80 GB. Export merged the selected adapter into the pinned base and then quantized the resulting BF16 GGUF.

The data covers a reviewed AuraGo snapshot, not every current tool or integration. Exact existing evaluation questions were kept out of correction training, but related synthetic templates overlap conceptually. Historical dialogs were already used during earlier Ling development and are not a pristine unseen benchmark.

Evaluation and limitations

Before quantization, the corrected adapter obtained:

Check	Result
Exact historical tool calls, including arguments	73/75
Parser/schema-valid historical assistant turns	147/147
Strict generated-history fixture chains	4/4
Exact calls in the new synthetic tool tests	12/12

These are limited contract tests with simulated tool results, not an overall answer-quality score. Exact matching may reject semantically equivalent arguments; nonempty prose is not proof of a correct answer. The full suite was not repeated on the quantized export.

The actual Q4_K_M file passed a local smoke test covering loading, authenticated completions, tokenizer samples, one native structured browser call with exact arguments, a follow-up after a simulated tool result, a German translation and completed streaming. The native server removes the API-only strict field from tool definitions; after that normalization, its rendered tool prompt matched the training template. No real tools were executed. This does not certify AuraGo integration, prompt-cache behavior, native GPU execution or long-context quality.

Known limitations remain:

• Two historical Netlify connectivity questions select get_site instead of check_connection.
• German summaries can omit facts or ignore sentence-count constraints; messages can copy instructions into their output.
• Arithmetic and explanations can be incorrect, including when the final number happens to be correct.
• Capability, memory and trust-boundary explanations can mix concepts or use unnatural wording.
• The model is not a substitute for runtime permissions, credential isolation, tool validation or review of consequential actions.

License

The base model and this derivative are distributed under Apache-2.0. The original license is included as LICENSE. This repository distributes the model artifact, not the AuraGo application or the inference-engine source.

Hashes and export provenance are recorded in provenance.json.