Changes / Feeds / unsloth/llama-3-8b-Instruct-bnb-4bit
unsloth/llama-3-8b-Instruct-bnb-4bit
11 changes to a tracked key since 2024-04-18, newest first. Head as last read: fd5a4dc328319c1cfe9489eccfb9c6406bdfd469.
| Date | Advisory | What changed, and what it does |
|---|---|---|
| 2024-08-07 | x_NXMC-2024-0233 breaking | config.json pad_token_id unset to 128255; the tokenizer's pad token, which training pads with, did not change (2024-08-07).158f943c2f39 -> 4864c68aa492 Open in app |
| 2024-08-07 | x_NXMC-2024-0232 | generation defaults changed (max_length 4096 to 8192); a call that does not pass these explicitly now generates differently (2024-08-07).a40f036948b6 -> 158f943c2f39 Open in app |
| 2024-08-07 | x_NXMC-2024-0231 breaking | hidden size 8192 to 4096, MLP width 28672 to 14336, attention heads 64 to 32, layer count 80 to 32; derived parameters 70.55B to 8.03B (-62.52B, -125.05 GB of bf16 weights) (2024-08-07).3c840fe7fe54 -> a40f036948b6 Open in app |
| 2024-08-06 | x_NXMC-2024-0230 breaking | quantization_config changed; the stored format of the weights is different (2024-08-06).c467eea304f1 -> 3c840fe7fe54 Open in app |
| 2024-08-06 | x_NXMC-2024-0229 breaking | generation_config.json pad_token_id unset to 128255; generate() fills finished sequences in a batch with this id, while training pads with the tokenizer's pad token, which did not change; generation defaults changed (do_sample unset to true, max_length unset to 4096, temperature unset to 0.6, top_p unset to 0.9); a call that does not pass these explicitly now generates differently (2024-08-06).7400af6f3bfa -> 7a83eed3a685 Open in app |
| 2024-08-06 | x_NXMC-2024-0228 breaking | hidden size 4096 to 8192, MLP width 14336 to 28672, attention heads 32 to 64, layer count 32 to 80; derived parameters 8.03B to 70.55B (+62.52B, +125.05 GB of bf16 weights); quantization_config changed; the stored format of the weights is different (2024-08-06).2950abc9d0b3 -> 7400af6f3bfa Open in app |
| 2024-05-16 | x_NXMC-2024-0227 | config.json eos_token_id 128001 to 128009, but generation_config.json still declares [128001,128009], which generate() reads first, so default generation stops on the same tokens; only code that reads config.eos_token_id directly sees the new value (2024-05-16).04fda3e5bcfd -> 2950abc9d0b3 Open in app |
| 2024-05-15 | x_NXMC-2024-0226 breaking | tokenizer_config.json is still in the repository but is no longer valid JSON, so none of its settings can be read from it (2024-05-15).89e4fd4e68bf -> 2bc2f21daa60 Open in app |
| 2024-05-05 | x_NXMC-2024-0225 | pad token "<|end_of_text|>" to "<|reserved_special_token_250|>"; batched training masks a different id; tokenizer model_max_length 8192 to 1e+30; truncation defaults change (2024-05-05).d8d00d975d40 -> 89e4fd4e68bf Open in app |
| 2024-04-23 | x_NXMC-2024-0224 breaking | EOS stop list 128001 to [128001,128009]; generation now also stops on 128009, and still stops on every token it stopped on before (2024-04-23).3fe39f383d04 -> d8d00d975d40 Open in app |
| 2024-04-19 | x_NXMC-2024-0223 | pad token unset to "<|end_of_text|>"; batched training masks a different id; the chat template text changed but rendered byte-identically on the three fixed conversations (system+user, multi-turn, tool call), all of which end on a user turn; a conversation ending on an assistant turn (training data, prefill) was not rendered (2024-04-19).83226572bdb1 -> 1604b9078935 Open in app |
Pin it, and let a bot move the pin
from_pretrained("unsloth/llama-3-8b-Instruct-bnb-4bit", revision="fd5a4dc328319c1cfe9489eccfb9c6406bdfd469")With the Renovate preset, a pull request moves this pin when the head moves, and links back here.
[](https://www.neurarch.com/changes-feeds/models/unsloth/llama-3-8b-Instruct-bnb-4bit.html)