Model · baize-v2-7b
Baize v2 7B
Developer: Project Baize (UC San Diego; Sun Yat-sen University)
Availability: available · checked 2026-09-24 · source
Raw record: /data/models/baize-v2-7b.json
Fields
- id
- baize-v2-7b
- identifiers
- huggingface
- project-baize/baize-v2-7b
- developer
- Project Baize (UC San Diego; Sun Yat-sen University)
- release_date
- 2023-05-23recorded · sourcenote: Project README, dated entry: '[May 23, 2023] We are releasing Baize v2! Check out the 7B and 13B model.' HF repo creation date matches.
- weights_status
- open
- availability
- available · checked 2026-09-24 · source
- license
- cc-by-nc-4.0partial · sourcenote: Card metadata only (`license: cc-by-nc-4.0`); the card carries no separate license text. The weights are merged with LLaMA, whose own terms are not addressed on the card.
- architecture
- family
- decoder_only
- note
- family read from config 'architectures': ['LlamaForCausalLM'].
- n_layers
- 32recorded · source
- hidden_size
- 4096recorded · source
- n_heads
- 32recorded · source
- vocab_size
- 32000recorded · source
- context_length
- 2048recorded · source
- positional_encoding
- rotary (RoPE)recorded · sourcenote: Propagated from llama-7b: the card says the LoRA-tuned checkpoint 'has been merged with LLaMA' and neither card nor paper describes an architecture change.
- training_data
- Baize self-chat dialogues (ChatGPT chatting with itself) plus Alpaca data, per the project README for the Baize family; the v2 card says only that the model was 'trained with supervised fine-tuning (SFT) and self-distillation with feedback (SDF)'. The exact v2 data mixture is not stated.partial · sourcenote: README: 'It uses 100k dialogs generated by letting ChatGPT chat with itself. We also use Alpaca's data to improve its performance.' The README's own training command (alpaca,stackoverflow,quora) is for the v1 sizes.
- techniques
- primary_sources
- record_history
- date:2026-09-24 · change:ingested as candidate from HF (project-baize/baize-v2-7b@e4731c2c2671e2d0b47b5eba08c753ca21671fab) · by:ingest_hf.py ·date:2026-09-24 · change:preparer: developer, release date (README), positional encoding (propagated from llama-7b), training data (partial), 3 edges prepared + 1 PANEL-NEEDED; sources: card, README, arXiv 2304.01196 · by:claude (preparer, Sonnet 5) ·date:2026-09-24 · change:ruling: Wilson, after 3-tier panel split (no majority), see session log: baize-sdf stub; trained_on baize-sdf added, direct feedback_from chatgpt rejected · by:claude (preparer, Sonnet 5) ·date:2026-09-24 · change:reviewed and promoted from staging (3 edge(s) accepted) · by:Wilson Pruitt ·
Parents
Weights descend
- fine_tuned_from → LLaMA 7B declared source Card: 'Baize is an open-source chat model fine-tuned with LoRA. This model is a 7B Baize-v2 ... This checkpoint has been merged with LLaMA so it's ready for use.' The LoRA is merged into LLaMA, so the weights descend from LLaMA. Size (7B, so llama-7b) is from the model's name; the README's v1 merge example uses base huggyllama/llama-7b. Intermediate step not recorded: the Baize paper (arXiv 2304.01196, Table 3) has v2 built on Baize v1.5, itself LLaMA-7B fine-tuned, and v1.5 is not recorded in Stemma. This edge is true but skips that step (Wilson's ruling, 2026-09-24: keep, note the gap).
Training data
- trained_on → Baize self-chat data declared source Card: trained on self-chat data (SFT + SDF); paper: 'leveraging ChatGPT to engage in a conversation with itself'. The `baize` dataset record describes the v1 corpus; the README ships separate v1 and v2 collection scripts (collect.py, collect_v2.py), so v2's exact dialogues are not pinned down. Reviewer: confirm the family-level edge is acceptable.
- trained_on → Baize SDF data (ChatGPT-ranked self-generations) declared source Card: v2 trained with 'self-distillation with feedback (SDF)'. Paper Table 3 lists Baize-v2-7B as SDF applied to Baize-v1.5-7B: 'we apply new LoRA modules to all linear layers in Baize v1.5'. SDF data is Baize v1.5's own generations for Quora instructions, ranked by ChatGPT.
Children
No edges recorded.
Read in
No station on the reading path has touched this record yet.