Instruct-thinking hybrid tuning
According to the model card, Unsloth tuning for 3 epochs on the TeichAI Claude high-reasoning dataset created an instruct and thinking hybrid without updating core knowledge.
Open Source Model Profile · DavidAU
Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-Reasoning
Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-Reasoning is an 8.03B-parameter Llama instruct-thinking hybrid from DavidAU. According to the model card, it was tuned with Unsloth for 3 epochs.
- Publisher
- DavidAU
- Task
- text-generation
- Model type
- llama
- License
- apache-2.0
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-Reasoning is published by DavidAU as a Llama text-generation fine-tune. The captured configuration identifies LlamaForCausalLM with model type llama, and Safetensors metadata reports 8,030,261,248 parameters. According to the model card, Unsloth tuning for 3 epochs produced an instruct and thinking hybrid from allura-forge/Llama-3.3-8B-Instruct, with card data recording apache-2.0.
Recorded capabilities
Self-generating thinking tags
According to the model card, no system prompt is needed because thinking tags self-generate, with hub tags recording thinking, reasoning, and instruct entries.
Chat and roleplay sampler notes
According to the model card, the publisher documents a 1.5 smoothing factor for KoboldCpp, text-generation-webui, and Silly Tavern chat or roleplay use.
8B Llama Transformers record
The captured configuration identifies LlamaForCausalLM with model type llama and about 8.03B Safetensors parameters with Transformers support.
Use cases in the source record
- Thinking-style text-generation experiments using the documented self-generating thinking-tag behavior without a system prompt.
- Chat and roleplay trials following the publisher-documented 1.5 smoothing-factor sampler guidance.
- Transformers-based instruction experiments using the captured 8B Llama configuration and Safetensors weights.
Limitations and unknowns
- According to the model card, the publisher describes an unreleased-source provenance account; Ethen has not independently verified that lineage narrative.
- Much of the captured card is a sample orbital-mechanics generation, which is an output illustration rather than measured capability evidence.
- No evaluation results or independently verified context-window value were extracted.
- Provider state is historical snapshot data and should be refreshed before being presented as current.
Source and provenance
Source: DavidAU/Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-Reasoning
Captured: Unknown. Processed: 2026-09-07T19:35:39.015150+00:00.
Llama3.3-8B-Instruct-Thinking-Claude-4.5-Opus-High-Reasoning What madness is this? Someone found "Llama3.3-8B" source (never publicly released) in the "wild", then it was adjusted back to 128k and then I added my own special madness: Training the model with Unsloth (3 epochs) and Claude 4.5-Opus High Reasoning dataset. This has created an Instruct/Thinking hybrid (128k context, Llama 3.3 model). Note this tuning was only to create an instruct/thinking model, not to update the model's core knowledge / root training. 1 example at bottom of the page. HERETIC / Uncensored Version: https://huggingface.co/DavidAU/Llama3.3-8B-Instruct-Thin…
F001F002F003F004F005F006F007F008F009F010F012F013F014F015F016F017F019F020F021F022