schneewolflabs/A3-Instruct
Image-Text-to-Text • 13B • Updated • 8
id string | layer string | source string | messages string | image image | reasoning string |
|---|---|---|---|---|---|
L1-liuhaotian_llava-instruct-150k-00056412 | L1 | liuhaotian/LLaVA-Instruct-150K | [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "What could be the potential negative consequences of the child eating cake in this manner?"}]}, {"role": "assistant", "content": "While the child in the image is enjoying their cake with frosting and cake all over their hands and face, there are... | null | |
L1-huggingfacem4_the_cauldron@textvqa-00221801 | L1 | HuggingFaceM4/the_cauldron@textvqa | [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "What word means fast on the poster?\nYour answer should be very brief."}]}, {"role": "assistant", "content": "Zoom."}] | null | |
L1-liuhaotian_llava-instruct-150k-00056739 | L1 | liuhaotian/LLaVA-Instruct-150K | [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "What safety concerns might arise from this scenario?"}]}, {"role": "assistant", "content": "Performing skate tricks in the dark, especially at night in a parking lot, can pose several safety concerns. First, the limited visibility in the dark in... | null | |
L4-jackrong_competitive-programming-python--00013151 | L4 | Jackrong/Competitive-Programming-python-blend | "[{\"role\": \"user\", \"content\": \"You are given a string $$$s$$$, consisting of lowercase Latin (...TRUNCATED) | Not supported with pagination yet | null |
L4-schneewolflabs_i-dpo-00000215-r24 | L4 | schneewolflabs/i-DPO | "[{\"role\": \"system\", \"content\": \"You are a concise, helpful assistant.\"}, {\"role\": \"user\(...TRUNCATED) | Not supported with pagination yet | null |
L4-a2-tool-orpo-v1_jsonl-00002279-r9 | L4 | A2-tool-orpo-v1.jsonl | "[{\"role\": \"system\", \"content\": \"# Tools\\n\\nYou have access to the following functions:\\n\(...TRUNCATED) | Not supported with pagination yet | None of the available tools cover this, so I'll say I can't rather than force a call. |
L4-schneewolflabs_i-dpo-00000559-r17 | L4 | schneewolflabs/i-DPO | "[{\"role\": \"system\", \"content\": \"You are Luna, an AI assistant developed by Schneewolf Labs.\(...TRUNCATED) | Not supported with pagination yet | null |
L1-liuhaotian_llava-instruct-150k-00097340 | L1 | liuhaotian/LLaVA-Instruct-150K | "[{\"role\": \"user\", \"content\": [{\"type\": \"image\"}, {\"type\": \"text\", \"text\": \"What is(...TRUNCATED) | null | |
L1-liuhaotian_llava-instruct-150k-00117131 | L1 | liuhaotian/LLaVA-Instruct-150K | "[{\"role\": \"user\", \"content\": [{\"type\": \"image\"}, {\"type\": \"text\", \"text\": \"Is the (...TRUNCATED) | null | |
L4-jackrong_logicmind-chat-reasoning-sft-30-00000842 | L4 | Jackrong/LogicMind-Chat-Reasoning-SFT-300K | "[{\"role\": \"user\", \"content\": \"In an alternate timeline, Ferdinand Magellan and his expeditio(...TRUNCATED) | Not supported with pagination yet | "Okay, so I need to help expand on this alternate timeline scenario where Magellan encounters a mix (...TRUNCATED) |
End of preview. Expand in Data Studio
ArtemisMix-v1.1
The as-trained Stage-2 corpus for Artemis1-13B (Schneewolf Labs' Qwen3-VL-ViT-grafted-onto-A2/Mistral vision-language flagship).
Extends schneewolflabs/ArtemisMix-v1
with 14,816 non-reasoning creative-writing rows derived from
schneewolflabs/Athanorlite-DPO
(DPO collapsed to SFT, chosen only, bare assistant content so the A2
chat template renders an empty <think></think> block — i.e. thinking-off
mode). The bucket balances v1's reasoning-heavy L4 with a direct-response
creative gear so Artemis1 doesn't compulsively reason on every prompt.
Composition
| Layer | Rows | Share | Sources |
|---|---|---|---|
| L1 general multimodal instruction | 150,000 | 41.1% | LLaVA-Instruct 120K + the_cauldron VQA 30K |
| L4 text rehearsal | 214,816 | 58.9% | v1's L4 (200,000) + Athanorlite-DPO (14,816) |
| Total | 364,816 | 100% |
The anti-drift bedrock (i-DPO + A2-tool-orpo, oversampled in v1) still
provides ~16% of the corpus; the new Athanorlite rows add ~4% of explicit
non-reasoning text so the model learns the empty-<think> mode.
Delta from v1
Only the new 14,816 rows differ from v1. Schema is identical:
| field | type | notes |
|---|---|---|
id |
string | L4-athanorlite-{idx} for the new rows |
layer |
string | L4 |
source |
string | schneewolflabs/Athanorlite-DPO |
messages |
string | JSON-serialized chat: system? + user(prompt) + assistant(chosen) — bare content (no <think>) |
image |
Image | null (text-only) |
reasoning |
string | null |
messages is JSON-encoded for Arrow compatibility — json.loads at load time.
Why empty-<think>?
A2's chat template auto-wraps the final assistant turn with <think>{reasoning}</think>\n\n{answer}.
If content contains a <think> block, the template renders it as-is; if
not (bare content), reasoning_content is empty and the template emits
<think>\n\n</think>\n\n{answer}. Athanorlite rows ship bare assistant
content, so they teach the model the empty-<think> (thinking-off) mode.
The reasoning-tagged rows from v1 keep their real <think> blocks. Result:
a hybrid model where enable_thinking at inference toggles between modes.
Lineage
schneewolflabs/ArtemisMix-v1— predecessor; clean public-data L1+L4 slice (350,000 rows). Use this if you want the standalone reasoning + multimodal mix without the creative bucket.schneewolflabs/Athanorlite-DPO— upstream creative-writing DPO pairs.
License
apache-2.0 umbrella, consistent with the A1/A2/A3 lineage. See v1 for the constituent source licenses (mostly apache-2.0 / mit / cc-by; Jackrong reasoning sets and competitive-programming blend are mixed-permissive).
- Downloads last month
- 137
Models trained or fine-tuned on schneewolflabs/ArtemisMix-v1.1