ApollonLabs/Heliactis-1-4B-GGUF

Heliactis-1-4B — GGUF

Heliactis means "sunbeam". It is the ancient-style form of the living Greek word ηλιαχτίδα (ηλιακτίδα), built from the ancient hḗlios ("sun") and aktís ("ray"). The Heliactis models are the rays; the larger Helios models are the sun itself. The lab is named after Apollo (Apollon), the Greek god of light.

Quantized builds of Heliactis 1, an Apollon Labs fine-tune of XHToken/Spark-X2.5-4B, for llama.cpp, LM Studio, Ollama and anything else that reads GGUF.

What it is for: Greek that does not fall apart, code that stays intact, and tool calling that knows when to stay quiet.

Files

File	Size	Gate
Heliactis-1-4B-F16.gguf	7.66 GiB	passed
Heliactis-1-4B-Q8_0.gguf	4.07 GiB	passed
Heliactis-1-4B-Q6_K.gguf	3.15 GiB	passed (imatrix)
Heliactis-1-4B-Q5_K_M.gguf	2.77 GiB	passed (imatrix)
Heliactis-1-4B-Q4_K_M.gguf	2.42 GiB	passed (imatrix) · the default, runs on 6 GB VRAM or plain CPU

Every file listed here went through the full gate below. Nothing is shipped unmeasured.

Usage

llama-server -m Heliactis-1-4B-Q4_K_M.gguf --jinja -c 8192 \
             --cache-type-k q8_0 --cache-type-v q8_0

Context window: 1,048,576 tokens, inherited unchanged from the base (max_position_embeddings; 27 of 36 layers use a 512-token sliding window, which keeps the KV cache small). -c 8192 above is only a default that fits small GPUs: raise it as far as your memory allows. The KV cache at q8_0 costs 14,976 bytes per token, about 14.6 GiB for the full window. Long-context recall was not re-measured for this release; see the limitations below.

The chat template is embedded in the GGUF. The model was trained and evaluated with thinking disabled. Call it that way. --jinja is required for tool calling, otherwise the toolbox never reaches the model.

Identity

This model was not trained on any self-identity data. If identity matters, set it in the system prompt:

You are Heliactis 1, a Greek/English assistant built by Apollon Labs
on top of Spark-X2.5-4B.

A base-model defect, reduced

On Spark-X2.5-4B, 502 vocabulary entries begin with an orphan UTF-8 byte. When the model emits one, the output breaks and llama-server answers HTTP 500, even on an ordinary translation prompt. We reported it upstream. Heliactis 1 was trained against those tokens: at the worst position we measured, their probability fell from 12.7% (our unreleased intermediate fine-tune, the starting point of this training) to 0.17%. Rarer, not impossible.

Block them at inference to close this route (ban-ids.json ships in this repo):

llama-server -m Heliactis-1-4B-Q4_K_M.gguf --jinja \
  $(python -c "import json;print(' '.join(f'-l {i}-inf' for i in json.load(open('ban-ids.json'))))")

Via the API, per request: "logit_bias": [[<id>, false], ...] for the same ids. A system prompt does not fix it. Full reproduction, training trade-off and a second route that blocking does not close: DEFECT.md.

Measured results

Instrument: 12-part gate, thinking disabled, --jinja, KV q8_0, -c 8192. Each value comes from the gate's own results files.

Test	Q4_K_M	Q5_K_M	Q6_K	Q8_0	F16	bar
Tool abstention	15/15	15/15	15/15	15/15	15/15	≥ 9
Overall tool score	45/45	45/45	45/45	45/45	45/45	≥ 28
Tool calls, Greek	10/10	10/10	10/10	10/10	10/10	≥ 7
Tool calls, English	20/20	20/20	20/20	20/20	20/20	≥ 17.5
HumanEval pass@1	0.7439	0.7622	0.7805	0.7622	0.7866	≥ 0.6952
Code fencing	60/60	57/60	58/60	59/60	58/60	≥ 53
Capability eval	52/60	53/60	55/60	54/60	54/60	≥ 51
Greek character validity	0.9996	0.9983	0.9974	0.9974	0.9968	≥ 0.9943
Greek vocabulary	0.9553	0.9536	0.9528	0.9549	0.9568	≥ 0.949
Greek meaning	0.76	0.79	0.78	0.79	0.80	≥ 0.70
Median answer length (tokens)	242	251.5	256.5	253.5	254.5	≤ 266
Truncated answers	3	5	2	2	3	≤ 9

Honest limitations

• The orphan-byte defect is reduced, not removed. See above; block the ids to close this route, and see the Lao note below.
• Do not use this model for Lao. A second route to the same broken output exists: the model can emit a token that ends mid-character and then not finish it. Blocking the 502 ids does not stop this. In our test (greedy, one prompt per language) a Lao paragraph ended in HTTP 500 on this model and on our unreleased intermediate fine-tune, but not on the base model; Thai passed on all three.
• Lao and Thai got worse. The 502 blocked tokens are pieces of those scripts. Perplexity vs our unreleased intermediate fine-tune, fp32, 5 short outside texts (151 tokens, so the noise is large): English 17.09 → 16.81 · code 18.44 → 19.30 · Greek 21.13 → 27.13 · mixed languages 56.73 → 67.90 · Lao/Thai 56.49 → 78.85. The Greek line is one sentence; the gate's Greek vocabulary test is the measure we trust, and it passed.
• Greek meaning (0.76 on Q4_K_M) has a wide ±0.08 band. It detects collapse, not small changes.
• Abstention was measured on N=15 prompts. Evidence, not proof.
• Long context was not re-measured for this release. The architecture and KV cost are the base model's (14,976 bytes/token at q8_0; 9 of 36 layers full attention). An unreleased earlier fine-tune on the same base recalled 3/3 two-hop facts at 231k tokens. We do not carry that number over as a claim for this model.
• 24 of the 554 code training examples use Greek identifiers. Known, not yet fixed.

Training

LoRA (r=16, alpha=32). Stage 1: SFT on 2014 examples (1212 Greek, 554 code, 248 tool-calling; 742 of the 866 toolbox rows are decoys that teach abstention). Dataset sha256: cd58f73ad7fbb339de87a3bef22af6c2ff996b971f829978c78b4f70703f26e6. Stage 2: on-policy unlikelihood on the 502 tokens (856 of the model's own answers, weight 10), with the same SFT set alongside to hold the rest in place. imatrix: 100 × 512-token chunks of Greek text.

License and attribution

Apache 2.0, inherited from the base model XHToken/Spark-X2.5-4B. Derivative work by Apollon Labs.

---

Apollon Labs — Towards the Sun.