Mistral-7B-Instruct-v0.2 is explicitly documented as an instruction-tuned version of the Mistral-7B-v0.2 base model. The architecture is a decoder-only transformer with 32 layers, 4096 hidden dimensions, and 14336 intermediate dimensions. It utilizes Grouped-Query Attention (GQA) and Rotary Position Embeddings (RoPE) with a theta value of 1e6. Notably, this version (v0.2) removes the Sliding-Window Attention (SWA) found in v0.1 to support a full 32k context window. While the high-level architecture is well-documented in the 'Mistral 7B' paper and release notes, the specific fine-tuning methodology for the 'Instruct' variant is described generally as 'two-stage instruction tuning' without exhaustive procedural detail.

Data transparency is a significant weakness. Mistral AI states the model was fine-tuned on 'publicly available conversation datasets' but does not provide a specific list, proportions, or a breakdown of the pretraining or instruction data. There is no public documentation on data filtering, cleaning, or the specific mix of code, web, and book data used. The company explicitly refuses to disclose detailed data sources for competitive reasons, which falls under the 'vague marketing claims' category in the scoring guidelines.

The model uses a Byte-fallback BPE tokenizer with a vocabulary size of 32,000. The tokenizer is publicly accessible via the 'mistral-common' library and Hugging Face 'transformers'. Documentation confirms it handles a wide range of inputs by falling back to bytes for out-of-vocabulary characters. The special tokens for instruction formatting ([INST], [/INST]) are clearly defined and documented for developers to ensure alignment between training and inference.

The model is a dense architecture with a clearly stated total of 7.3 billion parameters. Unlike MoE models, all parameters are active during inference. Detailed architectural specifications (layer counts, head counts, and dimension sizes) are available in the model configuration files. However, there is no public documentation on the specific parameter distribution across attention vs. feed-forward networks beyond what can be inferred from the standard transformer block structure.

Compute transparency is nearly non-existent. While it is known that the model was trained on a CoreWeave cluster, Mistral AI has not disclosed the total GPU hours, hardware counts (e.g., number of H100s), training duration, or the carbon footprint. Third-party estimates exist, but official verifiable data is absent, leading to a low score based on the 'no compute disclosure' red flag.

Mistral AI provides benchmark results for MMLU, GSM8K, and others in their blog posts and paper, but they do not release the exact evaluation code or the specific prompts/few-shot examples used to achieve those scores. While the model is frequently tested by third parties on leaderboards like LMSYS, the lack of official reproduction instructions and prompt transparency prevents a higher score. (Score adjusted for disclosed external findings regarding benchmark performance consistency).

The model generally identifies itself correctly as an AI developed by Mistral AI and is aware of its versioning (v0.2). It does not typically claim to be a competitor's model (like GPT-4). However, early documentation for v0.2 was initially inconsistent regarding its base model (v0.1 vs v0.2), which was later corrected. It maintains a coherent identity across standard deployments.

The model is released under the Apache 2.0 license, which is a highly permissive, standard open-source license. The terms are clear, allowing for commercial use, modification, and distribution without significant restrictions. There are no known conflicting terms between the weights and the reference code provided by Mistral AI.

VRAM requirements are well-documented by the community and supported by official configuration files. At FP16, the model requires ~14-15GB of VRAM, while 4-bit quantization (Q4) reduces this to ~4-5GB, making it accessible on consumer hardware. Memory scaling for the 32k context window is understood, though official documentation on the specific accuracy-tradeoffs of different quantization levels is primarily provided by third-party contributors rather than Mistral AI directly.

Mistral uses semantic-like versioning (v0.1, v0.2, v0.3), which is a positive practice. However, the transition from v0.1 to v0.2 involved silent updates to the model card and documentation regarding the base model and context window features. There is no detailed, centralized changelog or formal deprecation policy for older versions, making it difficult for developers to track subtle behavioral changes between minor releases.