Kolibri
Collection
A series of English/Russian instruction-following models.
•
6 items
•
Updated
•
1
This is an instruction following model (based on Qwen2.5-7B base) optimized for Russian language.
The model was trained in two phases: SFT (training data composition is similar to kolibri-mistral-0427) and RLHF.
Current RLHF pipeline leads to degradation on IFEval, but the overall 'vibe' of the model improves significantly. I am currently investigating the causes of this degradation and exploring methods to further enhance instruction-following capabilities.
The model uses ChatML template. Adding a system prompt will likely improve the model's performance on your tasks (experiment with it).
Instruction following evals
The model was tested using the following benchmarks:
| Eval name | Strict Value | Loose Value |
|---|---|---|
| Avg. | 43.00 | 49.17 |
| ifeval-prompt-level | 38.63 | 46.21 |
| ifeval-instruction-level | 51.20 | 57.5 |
| ru-ifeval-prompt-level | 35.30 | 40.48 |
| ru-ifeval-instruction-level | 46.88 | 52.52 |
Russian LLM Arena (proxy eval via JINA)
The table below approximates Russian LLM Arena scores using the JINA Judge model. Take it with a grain of salt.
| Model Name | Score | 95% CI | Avg Tokens |
|---|---|---|---|
| gpt-4-1106-preview | 82.8 | (-2.8, 2.6) | 541 |
| gpt-4o-mini | 75.3 | (-2.2, 2.8) | 448 |
| qwen-2.5-72b-it | 73.1 | (-3.0, 3.1) | 557 |
| gemma-2-9b-it-sppo-iter3 | 70.6 | (-3.7, 3.0) | 509 |
| gemma-2-27b-it | 68.7 | (-2.9, 3.8) | 472 |
| t-lite-instruct-0.1 | 67.5 | (-4.2, 2.7) | 810 |
| gemma-2-9b-it | 67.0 | (-3.0, 3.8) | 459 |
| suzume-llama-3-8B-multilingual-orpo-borda-half | 62.4 | (-3.0, 3.3) | 682 |
| glm-4-9b-chat | 61.5 | (-3.9, 3.3) | 568 |
| phi-3-medium-4k-instruct | 60.4 | (-3.8, 3.6) | 566 |
| sfr-iterative-dpo-llama-3-8b-r | 57.2 | (-3.8, 4.0) | 516 |
| kolibri-qwen2.5-7b-060225-rlhf-1 | 55.4 | (-3.1, 4.4) | 383 |
| c4ai-command-r-v01 | 55.0 | (-3.7, 4.4) | 529 |
| suzume-llama-3-8b-multilingual | 51.9 | (-3.1, 3.4) | 641 |
| mistral-nemo-instruct-2407 | 51.9 | (-3.0, 3.0) | 403 |
| yandex_gpt_pro | 50.3 | (-3.5, 3.0) | 345 |
| gpt-3.5-turbo-0125 | 50.0 | (0.0, 0.0) | 220 |
| hermes-2-theta-llama-3-8b | 49.3 | (-3.2, 3.7) | 485 |
| starling-lm-7b-beta | 48.3 | (-3.7, 3.9) | 629 |
| llama-3-8b-saiga-suzume-ties | 47.9 | (-3.9, 5.0) | 763 |
| llama-3-smaug-8b | 47.6 | (-4.3, 2.9) | 524 |
| vikhr-it-5.4-fp16-orpo-v2 | 46.8 | (-2.4, 2.2) | 379 |
| aya-23-8b | 46.1 | (-3.3, 3.6) | 554 |
| saiga_llama3_8b_v6 | 44.8 | (-2.9, 3.2) | 471 |
| qwen2-7b-instruct | 43.6 | (-3.5, 3.0) | 340 |
| vikhr-it-5.2-fp16-cp | 43.6 | (-3.6, 3.3) | 543 |
| openchat-3.5-0106 | 42.8 | (-2.5, 3.8) | 492 |
| kolibri-mistral-0427-upd | 42.3 | (-4.1, 4.0) | 551 |
| paralex-llama-3-8b-sft | 41.8 | (-3.7, 3.9) | 688 |
| llama-3-instruct-8b-sppo-iter3 | 41.7 | (-4.0, 3.6) | 502 |
| gpt-3.5-turbo-1106 | 41.5 | (-2.7, 2.5) | 191 |
| mistral-7b-instruct-v0.3 | 41.1 | (-4.1, 2.9) | 469 |
| gigachat_pro | 40.9 | (-3.2, 2.8) | 294 |
| openchat-3.6-8b-20240522 | 39.1 | (-2.9, 3.8) | 428 |
| vikhr-it-5.3-fp16-32k | 38.8 | (-3.2, 3.3) | 519 |
| hermes-2-pro-llama-3-8b | 38.4 | (-3.9, 3.9) | 463 |
| kolibri-vikhr-mistral-0427 | 34.5 | (-2.9, 3.1) | 489 |
| vikhr-it-5.3-fp16 | 33.5 | (-3.0, 3.8) | 523 |
| llama-3-instruct-8b-simpo | 32.7 | (-3.2, 2.7) | 417 |
| meta-llama-3-8b-instruct | 32.1 | (-3.6, 4.2) | 450 |
| neural-chat-7b-v3-3 | 25.9 | (-3.1, 3.2) | 927 |
| gigachat_lite | 25.4 | (-3.5, 2.7) | 276 |
| snorkel-mistral-pairrm-dpo | 10.3 | (-2.3, 2.6) | 773 |
| storm-7b | 3.7 | (-1.9, 1.7) | 419 |
- Downloads last month
- 11
Inference Providers
NEW
This model is not currently available via any of the supported Inference Providers.
The model cannot be deployed to the HF Inference API:
The model has no library tag.